block/blk-mq.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Block multiqueue core code
   4  *
   5  * Copyright (C) 2013-2014 Jens Axboe
   6  * Copyright (C) 2013-2014 Christoph Hellwig
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/backing-dev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/kmemleak.h>
  14 #include <linux/mm.h>
  15 #include <linux/init.h>
  16 #include <linux/slab.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/smp.h>
  19 #include <linux/llist.h>
  20 #include <linux/list_sort.h>
  21 #include <linux/cpu.h>
  22 #include <linux/cache.h>
  23 #include <linux/sched/sysctl.h>
  24 #include <linux/sched/topology.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/delay.h>
  27 #include <linux/crash_dump.h>
  28 #include <linux/prefetch.h>
  29
  30 #include <trace/events/block.h>
  31
  32 #include <linux/blk-mq.h>
  33 #include <linux/t10-pi.h>
  34 #include "blk.h"
  35 #include "blk-mq.h"
  36 #include "blk-mq-debugfs.h"
  37 #include "blk-mq-tag.h"
  38 #include "blk-pm.h"
  39 #include "blk-stat.h"
  40 #include "blk-mq-sched.h"
  41 #include "blk-rq-qos.h"
  42
  43 static void blk_mq_poll_stats_start(struct request_queue *q);
  44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  45
  46 static int blk_mq_poll_stats_bkt(const struct request *rq)
  47 {
  48         int ddir, sectors, bucket;
  49
  50         ddir = rq_data_dir(rq);
  51         sectors = blk_rq_stats_sectors(rq);
  52
  53         bucket = ddir + 2 * ilog2(sectors);
  54
  55         if (bucket < 0)
  56                 return -1;
  57         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  58                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  59
  60         return bucket;
  61 }
  62
  63 /*
  64  * Check if any of the ctx, dispatch list or elevator
  65  * have pending work in this hardware queue.
  66  */
  67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  68 {
  69         return !list_empty_careful(&hctx->dispatch) ||
  70                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  71                         blk_mq_sched_has_work(hctx);
  72 }
  73
  74 /*
  75  * Mark this ctx as having pending work in this hardware queue
  76  */
  77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  78                                      struct blk_mq_ctx *ctx)
  79 {
  80         const int bit = ctx->index_hw[hctx->type];
  81
  82         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
  83                 sbitmap_set_bit(&hctx->ctx_map, bit);
  84 }
  85
  86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  87                                       struct blk_mq_ctx *ctx)
  88 {
  89         const int bit = ctx->index_hw[hctx->type];
  90
  91         sbitmap_clear_bit(&hctx->ctx_map, bit);
  92 }
  93
  94 struct mq_inflight {
  95         struct hd_struct *part;
  96         unsigned int *inflight;
  97 };
  98
  99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
 100                                   struct request *rq, void *priv,
 101                                   bool reserved)
 102 {
 103         struct mq_inflight *mi = priv;
 104
 105         /*
 106          * index[0] counts the specific partition that was asked for.
 107          */
 108         if (rq->part == mi->part)
 109                 mi->inflight[0]++;
 110
 111         return true;
 112 }
 113
 114 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
 115 {
 116         unsigned inflight[2];
 117         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 118
 119         inflight[0] = inflight[1] = 0;
 120         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 121
 122         return inflight[0];
 123 }
 124
 125 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
 126                                      struct request *rq, void *priv,
 127                                      bool reserved)
 128 {
 129         struct mq_inflight *mi = priv;
 130
 131         if (rq->part == mi->part)
 132                 mi->inflight[rq_data_dir(rq)]++;
 133
 134         return true;
 135 }
 136
 137 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 138                          unsigned int inflight[2])
 139 {
 140         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 141
 142         inflight[0] = inflight[1] = 0;
 143         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
 144 }
 145
 146 void blk_freeze_queue_start(struct request_queue *q)
 147 {
 148         mutex_lock(&q->mq_freeze_lock);
 149         if (++q->mq_freeze_depth == 1) {
 150                 percpu_ref_kill(&q->q_usage_counter);
 151                 mutex_unlock(&q->mq_freeze_lock);
 152                 if (queue_is_mq(q))
 153                         blk_mq_run_hw_queues(q, false);
 154         } else {
 155                 mutex_unlock(&q->mq_freeze_lock);
 156         }
 157 }
 158 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 159
 160 void blk_mq_freeze_queue_wait(struct request_queue *q)
 161 {
 162         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 163 }
 164 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 165
 166 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 167                                      unsigned long timeout)
 168 {
 169         return wait_event_timeout(q->mq_freeze_wq,
 170                                         percpu_ref_is_zero(&q->q_usage_counter),
 171                                         timeout);
 172 }
 173 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 174
 175 /*
 176  * Guarantee no request is in use, so we can change any data structure of
 177  * the queue afterward.
 178  */
 179 void blk_freeze_queue(struct request_queue *q)
 180 {
 181         /*
 182          * In the !blk_mq case we are only calling this to kill the
 183          * q_usage_counter, otherwise this increases the freeze depth
 184          * and waits for it to return to zero.  For this reason there is
 185          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 186          * exported to drivers as the only user for unfreeze is blk_mq.
 187          */
 188         blk_freeze_queue_start(q);
 189         blk_mq_freeze_queue_wait(q);
 190 }
 191
 192 void blk_mq_freeze_queue(struct request_queue *q)
 193 {
 194         /*
 195          * ...just an alias to keep freeze and unfreeze actions balanced
 196          * in the blk_mq_* namespace
 197          */
 198         blk_freeze_queue(q);
 199 }
 200 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 201
 202 void blk_mq_unfreeze_queue(struct request_queue *q)
 203 {
 204         mutex_lock(&q->mq_freeze_lock);
 205         q->mq_freeze_depth--;
 206         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 207         if (!q->mq_freeze_depth) {
 208                 percpu_ref_resurrect(&q->q_usage_counter);
 209                 wake_up_all(&q->mq_freeze_wq);
 210         }
 211         mutex_unlock(&q->mq_freeze_lock);
 212 }
 213 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 214
 215 /*
 216  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 217  * mpt3sas driver such that this function can be removed.
 218  */
 219 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 220 {
 221         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 222 }
 223 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 224
 225 /**
 226  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 227  * @q: request queue.
 228  *
 229  * Note: this function does not prevent that the struct request end_io()
 230  * callback function is invoked. Once this function is returned, we make
 231  * sure no dispatch can happen until the queue is unquiesced via
 232  * blk_mq_unquiesce_queue().
 233  */
 234 void blk_mq_quiesce_queue(struct request_queue *q)
 235 {
 236         struct blk_mq_hw_ctx *hctx;
 237         unsigned int i;
 238         bool rcu = false;
 239
 240         blk_mq_quiesce_queue_nowait(q);
 241
 242         queue_for_each_hw_ctx(q, hctx, i) {
 243                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 244                         synchronize_srcu(hctx->srcu);
 245                 else
 246                         rcu = true;
 247         }
 248         if (rcu)
 249                 synchronize_rcu();
 250 }
 251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 252
 253 /*
 254  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 255  * @q: request queue.
 256  *
 257  * This function recovers queue into the state before quiescing
 258  * which is done by blk_mq_quiesce_queue.
 259  */
 260 void blk_mq_unquiesce_queue(struct request_queue *q)
 261 {
 262         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 263
 264         /* dispatch requests which are inserted during quiescing */
 265         blk_mq_run_hw_queues(q, true);
 266 }
 267 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 268
 269 void blk_mq_wake_waiters(struct request_queue *q)
 270 {
 271         struct blk_mq_hw_ctx *hctx;
 272         unsigned int i;
 273
 274         queue_for_each_hw_ctx(q, hctx, i)
 275                 if (blk_mq_hw_queue_mapped(hctx))
 276                         blk_mq_tag_wakeup_all(hctx->tags, true);
 277 }
 278
 279 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 280 {
 281         return blk_mq_has_free_tags(hctx->tags);
 282 }
 283 EXPORT_SYMBOL(blk_mq_can_queue);
 284
 285 /*
 286  * Only need start/end time stamping if we have iostat or
 287  * blk stats enabled, or using an IO scheduler.
 288  */
 289 static inline bool blk_mq_need_time_stamp(struct request *rq)
 290 {
 291         return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
 292 }
 293
 294 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 295                 unsigned int tag, unsigned int op, u64 alloc_time_ns)
 296 {
 297         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 298         struct request *rq = tags->static_rqs[tag];
 299         req_flags_t rq_flags = 0;
 300
 301         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 302                 rq->tag = -1;
 303                 rq->internal_tag = tag;
 304         } else {
 305                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
 306                         rq_flags = RQF_MQ_INFLIGHT;
 307                         atomic_inc(&data->hctx->nr_active);
 308                 }
 309                 rq->tag = tag;
 310                 rq->internal_tag = -1;
 311                 data->hctx->tags->rqs[rq->tag] = rq;
 312         }
 313
 314         /* csd/requeue_work/fifo_time is initialized before use */
 315         rq->q = data->q;
 316         rq->mq_ctx = data->ctx;
 317         rq->mq_hctx = data->hctx;
 318         rq->rq_flags = rq_flags;
 319         rq->cmd_flags = op;
 320         if (data->flags & BLK_MQ_REQ_PREEMPT)
 321                 rq->rq_flags |= RQF_PREEMPT;
 322         if (blk_queue_io_stat(data->q))
 323                 rq->rq_flags |= RQF_IO_STAT;
 324         INIT_LIST_HEAD(&rq->queuelist);
 325         INIT_HLIST_NODE(&rq->hash);
 326         RB_CLEAR_NODE(&rq->rb_node);
 327         rq->rq_disk = NULL;
 328         rq->part = NULL;
 329 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
 330         rq->alloc_time_ns = alloc_time_ns;
 331 #endif
 332         if (blk_mq_need_time_stamp(rq))
 333                 rq->start_time_ns = ktime_get_ns();
 334         else
 335                 rq->start_time_ns = 0;
 336         rq->io_start_time_ns = 0;
 337         rq->stats_sectors = 0;
 338         rq->nr_phys_segments = 0;
 339 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 340         rq->nr_integrity_segments = 0;
 341 #endif
 342         /* tag was already set */
 343         rq->extra_len = 0;
 344         WRITE_ONCE(rq->deadline, 0);
 345
 346         rq->timeout = 0;
 347
 348         rq->end_io = NULL;
 349         rq->end_io_data = NULL;
 350
 351         data->ctx->rq_dispatched[op_is_sync(op)]++;
 352         refcount_set(&rq->ref, 1);
 353         return rq;
 354 }
 355
 356 static struct request *blk_mq_get_request(struct request_queue *q,
 357                                           struct bio *bio,
 358                                           struct blk_mq_alloc_data *data)
 359 {
 360         struct elevator_queue *e = q->elevator;
 361         struct request *rq;
 362         unsigned int tag;
 363         bool clear_ctx_on_error = false;
 364         u64 alloc_time_ns = 0;
 365
 366         blk_queue_enter_live(q);
 367
 368         /* alloc_time includes depth and tag waits */
 369         if (blk_queue_rq_alloc_time(q))
 370                 alloc_time_ns = ktime_get_ns();
 371
 372         data->q = q;
 373         if (likely(!data->ctx)) {
 374                 data->ctx = blk_mq_get_ctx(q);
 375                 clear_ctx_on_error = true;
 376         }
 377         if (likely(!data->hctx))
 378                 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
 379                                                 data->ctx);
 380         if (data->cmd_flags & REQ_NOWAIT)
 381                 data->flags |= BLK_MQ_REQ_NOWAIT;
 382
 383         if (e) {
 384                 data->flags |= BLK_MQ_REQ_INTERNAL;
 385
 386                 /*
 387                  * Flush requests are special and go directly to the
 388                  * dispatch list. Don't include reserved tags in the
 389                  * limiting, as it isn't useful.
 390                  */
 391                 if (!op_is_flush(data->cmd_flags) &&
 392                     e->type->ops.limit_depth &&
 393                     !(data->flags & BLK_MQ_REQ_RESERVED))
 394                         e->type->ops.limit_depth(data->cmd_flags, data);
 395         } else {
 396                 blk_mq_tag_busy(data->hctx);
 397         }
 398
 399         tag = blk_mq_get_tag(data);
 400         if (tag == BLK_MQ_TAG_FAIL) {
 401                 if (clear_ctx_on_error)
 402                         data->ctx = NULL;
 403                 blk_queue_exit(q);
 404                 return NULL;
 405         }
 406
 407         rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
 408         if (!op_is_flush(data->cmd_flags)) {
 409                 rq->elv.icq = NULL;
 410                 if (e && e->type->ops.prepare_request) {
 411                         if (e->type->icq_cache)
 412                                 blk_mq_sched_assign_ioc(rq);
 413
 414                         e->type->ops.prepare_request(rq, bio);
 415                         rq->rq_flags |= RQF_ELVPRIV;
 416                 }
 417         }
 418         data->hctx->queued++;
 419         return rq;
 420 }
 421
 422 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 423                 blk_mq_req_flags_t flags)
 424 {
 425         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 426         struct request *rq;
 427         int ret;
 428
 429         ret = blk_queue_enter(q, flags);
 430         if (ret)
 431                 return ERR_PTR(ret);
 432
 433         rq = blk_mq_get_request(q, NULL, &alloc_data);
 434         blk_queue_exit(q);
 435
 436         if (!rq)
 437                 return ERR_PTR(-EWOULDBLOCK);
 438
 439         rq->__data_len = 0;
 440         rq->__sector = (sector_t) -1;
 441         rq->bio = rq->biotail = NULL;
 442         return rq;
 443 }
 444 EXPORT_SYMBOL(blk_mq_alloc_request);
 445
 446 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 447         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 448 {
 449         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 450         struct request *rq;
 451         unsigned int cpu;
 452         int ret;
 453
 454         /*
 455          * If the tag allocator sleeps we could get an allocation for a
 456          * different hardware context.  No need to complicate the low level
 457          * allocator for this for the rare use case of a command tied to
 458          * a specific queue.
 459          */
 460         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 461                 return ERR_PTR(-EINVAL);
 462
 463         if (hctx_idx >= q->nr_hw_queues)
 464                 return ERR_PTR(-EIO);
 465
 466         ret = blk_queue_enter(q, flags);
 467         if (ret)
 468                 return ERR_PTR(ret);
 469
 470         /*
 471          * Check if the hardware context is actually mapped to anything.
 472          * If not tell the caller that it should skip this queue.
 473          */
 474         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 475         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 476                 blk_queue_exit(q);
 477                 return ERR_PTR(-EXDEV);
 478         }
 479         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 480         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 481
 482         rq = blk_mq_get_request(q, NULL, &alloc_data);
 483         blk_queue_exit(q);
 484
 485         if (!rq)
 486                 return ERR_PTR(-EWOULDBLOCK);
 487
 488         return rq;
 489 }
 490 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 491
 492 static void __blk_mq_free_request(struct request *rq)
 493 {
 494         struct request_queue *q = rq->q;
 495         struct blk_mq_ctx *ctx = rq->mq_ctx;
 496         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 497         const int sched_tag = rq->internal_tag;
 498
 499         blk_pm_mark_last_busy(rq);
 500         rq->mq_hctx = NULL;
 501         if (rq->tag != -1)
 502                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 503         if (sched_tag != -1)
 504                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 505         blk_mq_sched_restart(hctx);
 506         blk_queue_exit(q);
 507 }
 508
 509 void blk_mq_free_request(struct request *rq)
 510 {
 511         struct request_queue *q = rq->q;
 512         struct elevator_queue *e = q->elevator;
 513         struct blk_mq_ctx *ctx = rq->mq_ctx;
 514         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 515
 516         if (rq->rq_flags & RQF_ELVPRIV) {
 517                 if (e && e->type->ops.finish_request)
 518                         e->type->ops.finish_request(rq);
 519                 if (rq->elv.icq) {
 520                         put_io_context(rq->elv.icq->ioc);
 521                         rq->elv.icq = NULL;
 522                 }
 523         }
 524
 525         ctx->rq_completed[rq_is_sync(rq)]++;
 526         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 527                 atomic_dec(&hctx->nr_active);
 528
 529         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 530                 laptop_io_completion(q->backing_dev_info);
 531
 532         rq_qos_done(q, rq);
 533
 534         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 535         if (refcount_dec_and_test(&rq->ref))
 536                 __blk_mq_free_request(rq);
 537 }
 538 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 539
 540 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 541 {
 542         u64 now = 0;
 543
 544         if (blk_mq_need_time_stamp(rq))
 545                 now = ktime_get_ns();
 546
 547         if (rq->rq_flags & RQF_STATS) {
 548                 blk_mq_poll_stats_start(rq->q);
 549                 blk_stat_add(rq, now);
 550         }
 551
 552         if (rq->internal_tag != -1)
 553                 blk_mq_sched_completed_request(rq, now);
 554
 555         blk_account_io_done(rq, now);
 556
 557         if (rq->end_io) {
 558                 rq_qos_done(rq->q, rq);
 559                 rq->end_io(rq, error);
 560         } else {
 561                 blk_mq_free_request(rq);
 562         }
 563 }
 564 EXPORT_SYMBOL(__blk_mq_end_request);
 565
 566 void blk_mq_end_request(struct request *rq, blk_status_t error)
 567 {
 568         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 569                 BUG();
 570         __blk_mq_end_request(rq, error);
 571 }
 572 EXPORT_SYMBOL(blk_mq_end_request);
 573
 574 static void __blk_mq_complete_request_remote(void *data)
 575 {
 576         struct request *rq = data;
 577         struct request_queue *q = rq->q;
 578
 579         q->mq_ops->complete(rq);
 580 }
 581
 582 static void __blk_mq_complete_request(struct request *rq)
 583 {
 584         struct blk_mq_ctx *ctx = rq->mq_ctx;
 585         struct request_queue *q = rq->q;
 586         bool shared = false;
 587         int cpu;
 588
 589         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 590         /*
 591          * Most of single queue controllers, there is only one irq vector
 592          * for handling IO completion, and the only irq's affinity is set
 593          * as all possible CPUs. On most of ARCHs, this affinity means the
 594          * irq is handled on one specific CPU.
 595          *
 596          * So complete IO reqeust in softirq context in case of single queue
 597          * for not degrading IO performance by irqsoff latency.
 598          */
 599         if (q->nr_hw_queues == 1) {
 600                 __blk_complete_request(rq);
 601                 return;
 602         }
 603
 604         /*
 605          * For a polled request, always complete locallly, it's pointless
 606          * to redirect the completion.
 607          */
 608         if ((rq->cmd_flags & REQ_HIPRI) ||
 609             !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
 610                 q->mq_ops->complete(rq);
 611                 return;
 612         }
 613
 614         cpu = get_cpu();
 615         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
 616                 shared = cpus_share_cache(cpu, ctx->cpu);
 617
 618         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 619                 rq->csd.func = __blk_mq_complete_request_remote;
 620                 rq->csd.info = rq;
 621                 rq->csd.flags = 0;
 622                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 623         } else {
 624                 q->mq_ops->complete(rq);
 625         }
 626         put_cpu();
 627 }
 628
 629 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 630         __releases(hctx->srcu)
 631 {
 632         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 633                 rcu_read_unlock();
 634         else
 635                 srcu_read_unlock(hctx->srcu, srcu_idx);
 636 }
 637
 638 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 639         __acquires(hctx->srcu)
 640 {
 641         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 642                 /* shut up gcc false positive */
 643                 *srcu_idx = 0;
 644                 rcu_read_lock();
 645         } else
 646                 *srcu_idx = srcu_read_lock(hctx->srcu);
 647 }
 648
 649 /**
 650  * blk_mq_complete_request - end I/O on a request
 651  * @rq:         the request being processed
 652  *
 653  * Description:
 654  *      Ends all I/O on a request. It does not handle partial completions.
 655  *      The actual completion happens out-of-order, through a IPI handler.
 656  **/
 657 bool blk_mq_complete_request(struct request *rq)
 658 {
 659         if (unlikely(blk_should_fake_timeout(rq->q)))
 660                 return false;
 661         __blk_mq_complete_request(rq);
 662         return true;
 663 }
 664 EXPORT_SYMBOL(blk_mq_complete_request);
 665
 666 int blk_mq_request_started(struct request *rq)
 667 {
 668         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 669 }
 670 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 671
 672 int blk_mq_request_completed(struct request *rq)
 673 {
 674         return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE;
 675 }
 676 EXPORT_SYMBOL_GPL(blk_mq_request_completed);
 677
 678 void blk_mq_start_request(struct request *rq)
 679 {
 680         struct request_queue *q = rq->q;
 681
 682         trace_block_rq_issue(q, rq);
 683
 684         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 685                 rq->io_start_time_ns = ktime_get_ns();
 686                 rq->stats_sectors = blk_rq_sectors(rq);
 687                 rq->rq_flags |= RQF_STATS;
 688                 rq_qos_issue(q, rq);
 689         }
 690
 691         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 692
 693         blk_add_timer(rq);
 694         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 695
 696         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 697                 /*
 698                  * Make sure space for the drain appears.  We know we can do
 699                  * this because max_hw_segments has been adjusted to be one
 700                  * fewer than the device can handle.
 701                  */
 702                 rq->nr_phys_segments++;
 703         }
 704
 705 #ifdef CONFIG_BLK_DEV_INTEGRITY
 706         if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
 707                 q->integrity.profile->prepare_fn(rq);
 708 #endif
 709 }
 710 EXPORT_SYMBOL(blk_mq_start_request);
 711
 712 static void __blk_mq_requeue_request(struct request *rq)
 713 {
 714         struct request_queue *q = rq->q;
 715
 716         blk_mq_put_driver_tag(rq);
 717
 718         trace_block_rq_requeue(q, rq);
 719         rq_qos_requeue(q, rq);
 720
 721         if (blk_mq_request_started(rq)) {
 722                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 723                 rq->rq_flags &= ~RQF_TIMED_OUT;
 724                 if (q->dma_drain_size && blk_rq_bytes(rq))
 725                         rq->nr_phys_segments--;
 726         }
 727 }
 728
 729 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 730 {
 731         __blk_mq_requeue_request(rq);
 732
 733         /* this request will be re-inserted to io scheduler queue */
 734         blk_mq_sched_requeue_request(rq);
 735
 736         BUG_ON(!list_empty(&rq->queuelist));
 737         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 738 }
 739 EXPORT_SYMBOL(blk_mq_requeue_request);
 740
 741 static void blk_mq_requeue_work(struct work_struct *work)
 742 {
 743         struct request_queue *q =
 744                 container_of(work, struct request_queue, requeue_work.work);
 745         LIST_HEAD(rq_list);
 746         struct request *rq, *next;
 747
 748         spin_lock_irq(&q->requeue_lock);
 749         list_splice_init(&q->requeue_list, &rq_list);
 750         spin_unlock_irq(&q->requeue_lock);
 751
 752         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 753                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 754                         continue;
 755
 756                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 757                 list_del_init(&rq->queuelist);
 758                 /*
 759                  * If RQF_DONTPREP, rq has contained some driver specific
 760                  * data, so insert it to hctx dispatch list to avoid any
 761                  * merge.
 762                  */
 763                 if (rq->rq_flags & RQF_DONTPREP)
 764                         blk_mq_request_bypass_insert(rq, false);
 765                 else
 766                         blk_mq_sched_insert_request(rq, true, false, false);
 767         }
 768
 769         while (!list_empty(&rq_list)) {
 770                 rq = list_entry(rq_list.next, struct request, queuelist);
 771                 list_del_init(&rq->queuelist);
 772                 blk_mq_sched_insert_request(rq, false, false, false);
 773         }
 774
 775         blk_mq_run_hw_queues(q, false);
 776 }
 777
 778 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 779                                 bool kick_requeue_list)
 780 {
 781         struct request_queue *q = rq->q;
 782         unsigned long flags;
 783
 784         /*
 785          * We abuse this flag that is otherwise used by the I/O scheduler to
 786          * request head insertion from the workqueue.
 787          */
 788         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 789
 790         spin_lock_irqsave(&q->requeue_lock, flags);
 791         if (at_head) {
 792                 rq->rq_flags |= RQF_SOFTBARRIER;
 793                 list_add(&rq->queuelist, &q->requeue_list);
 794         } else {
 795                 list_add_tail(&rq->queuelist, &q->requeue_list);
 796         }
 797         spin_unlock_irqrestore(&q->requeue_lock, flags);
 798
 799         if (kick_requeue_list)
 800                 blk_mq_kick_requeue_list(q);
 801 }
 802
 803 void blk_mq_kick_requeue_list(struct request_queue *q)
 804 {
 805         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 806 }
 807 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 808
 809 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 810                                     unsigned long msecs)
 811 {
 812         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 813                                     msecs_to_jiffies(msecs));
 814 }
 815 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 816
 817 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 818 {
 819         if (tag < tags->nr_tags) {
 820                 prefetch(tags->rqs[tag]);
 821                 return tags->rqs[tag];
 822         }
 823
 824         return NULL;
 825 }
 826 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 827
 828 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 829                                void *priv, bool reserved)
 830 {
 831         /*
 832          * If we find a request that is inflight and the queue matches,
 833          * we know the queue is busy. Return false to stop the iteration.
 834          */
 835         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 836                 bool *busy = priv;
 837
 838                 *busy = true;
 839                 return false;
 840         }
 841
 842         return true;
 843 }
 844
 845 bool blk_mq_queue_inflight(struct request_queue *q)
 846 {
 847         bool busy = false;
 848
 849         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 850         return busy;
 851 }
 852 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 853
 854 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 855 {
 856         req->rq_flags |= RQF_TIMED_OUT;
 857         if (req->q->mq_ops->timeout) {
 858                 enum blk_eh_timer_return ret;
 859
 860                 ret = req->q->mq_ops->timeout(req, reserved);
 861                 if (ret == BLK_EH_DONE)
 862                         return;
 863                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 864         }
 865
 866         blk_add_timer(req);
 867 }
 868
 869 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 870 {
 871         unsigned long deadline;
 872
 873         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 874                 return false;
 875         if (rq->rq_flags & RQF_TIMED_OUT)
 876                 return false;
 877
 878         deadline = READ_ONCE(rq->deadline);
 879         if (time_after_eq(jiffies, deadline))
 880                 return true;
 881
 882         if (*next == 0)
 883                 *next = deadline;
 884         else if (time_after(*next, deadline))
 885                 *next = deadline;
 886         return false;
 887 }
 888
 889 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 890                 struct request *rq, void *priv, bool reserved)
 891 {
 892         unsigned long *next = priv;
 893
 894         /*
 895          * Just do a quick check if it is expired before locking the request in
 896          * so we're not unnecessarilly synchronizing across CPUs.
 897          */
 898         if (!blk_mq_req_expired(rq, next))
 899                 return true;
 900
 901         /*
 902          * We have reason to believe the request may be expired. Take a
 903          * reference on the request to lock this request lifetime into its
 904          * currently allocated context to prevent it from being reallocated in
 905          * the event the completion by-passes this timeout handler.
 906          *
 907          * If the reference was already released, then the driver beat the
 908          * timeout handler to posting a natural completion.
 909          */
 910         if (!refcount_inc_not_zero(&rq->ref))
 911                 return true;
 912
 913         /*
 914          * The request is now locked and cannot be reallocated underneath the
 915          * timeout handler's processing. Re-verify this exact request is truly
 916          * expired; if it is not expired, then the request was completed and
 917          * reallocated as a new request.
 918          */
 919         if (blk_mq_req_expired(rq, next))
 920                 blk_mq_rq_timed_out(rq, reserved);
 921         if (refcount_dec_and_test(&rq->ref))
 922                 __blk_mq_free_request(rq);
 923
 924         return true;
 925 }
 926
 927 static void blk_mq_timeout_work(struct work_struct *work)
 928 {
 929         struct request_queue *q =
 930                 container_of(work, struct request_queue, timeout_work);
 931         unsigned long next = 0;
 932         struct blk_mq_hw_ctx *hctx;
 933         int i;
 934
 935         /* A deadlock might occur if a request is stuck requiring a
 936          * timeout at the same time a queue freeze is waiting
 937          * completion, since the timeout code would not be able to
 938          * acquire the queue reference here.
 939          *
 940          * That's why we don't use blk_queue_enter here; instead, we use
 941          * percpu_ref_tryget directly, because we need to be able to
 942          * obtain a reference even in the short window between the queue
 943          * starting to freeze, by dropping the first reference in
 944          * blk_freeze_queue_start, and the moment the last request is
 945          * consumed, marked by the instant q_usage_counter reaches
 946          * zero.
 947          */
 948         if (!percpu_ref_tryget(&q->q_usage_counter))
 949                 return;
 950
 951         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 952
 953         if (next != 0) {
 954                 mod_timer(&q->timeout, next);
 955         } else {
 956                 /*
 957                  * Request timeouts are handled as a forward rolling timer. If
 958                  * we end up here it means that no requests are pending and
 959                  * also that no request has been pending for a while. Mark
 960                  * each hctx as idle.
 961                  */
 962                 queue_for_each_hw_ctx(q, hctx, i) {
 963                         /* the hctx may be unmapped, so check it here */
 964                         if (blk_mq_hw_queue_mapped(hctx))
 965                                 blk_mq_tag_idle(hctx);
 966                 }
 967         }
 968         blk_queue_exit(q);
 969 }
 970
 971 struct flush_busy_ctx_data {
 972         struct blk_mq_hw_ctx *hctx;
 973         struct list_head *list;
 974 };
 975
 976 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 977 {
 978         struct flush_busy_ctx_data *flush_data = data;
 979         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 980         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 981         enum hctx_type type = hctx->type;
 982
 983         spin_lock(&ctx->lock);
 984         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 985         sbitmap_clear_bit(sb, bitnr);
 986         spin_unlock(&ctx->lock);
 987         return true;
 988 }
 989
 990 /*
 991  * Process software queues that have been marked busy, splicing them
 992  * to the for-dispatch
 993  */
 994 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 995 {
 996         struct flush_busy_ctx_data data = {
 997                 .hctx = hctx,
 998                 .list = list,
 999         };
1000
1001         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1002 }
1003 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1004
1005 struct dispatch_rq_data {
1006         struct blk_mq_hw_ctx *hctx;
1007         struct request *rq;
1008 };
1009
1010 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1011                 void *data)
1012 {
1013         struct dispatch_rq_data *dispatch_data = data;
1014         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1015         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1016         enum hctx_type type = hctx->type;
1017
1018         spin_lock(&ctx->lock);
1019         if (!list_empty(&ctx->rq_lists[type])) {
1020                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1021                 list_del_init(&dispatch_data->rq->queuelist);
1022                 if (list_empty(&ctx->rq_lists[type]))
1023                         sbitmap_clear_bit(sb, bitnr);
1024         }
1025         spin_unlock(&ctx->lock);
1026
1027         return !dispatch_data->rq;
1028 }
1029
1030 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1031                                         struct blk_mq_ctx *start)
1032 {
1033         unsigned off = start ? start->index_hw[hctx->type] : 0;
1034         struct dispatch_rq_data data = {
1035                 .hctx = hctx,
1036                 .rq   = NULL,
1037         };
1038
1039         __sbitmap_for_each_set(&hctx->ctx_map, off,
1040                                dispatch_rq_from_ctx, &data);
1041
1042         return data.rq;
1043 }
1044
1045 static inline unsigned int queued_to_index(unsigned int queued)
1046 {
1047         if (!queued)
1048                 return 0;
1049
1050         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1051 }
1052
1053 bool blk_mq_get_driver_tag(struct request *rq)
1054 {
1055         struct blk_mq_alloc_data data = {
1056                 .q = rq->q,
1057                 .hctx = rq->mq_hctx,
1058                 .flags = BLK_MQ_REQ_NOWAIT,
1059                 .cmd_flags = rq->cmd_flags,
1060         };
1061         bool shared;
1062
1063         if (rq->tag != -1)
1064                 goto done;
1065
1066         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1067                 data.flags |= BLK_MQ_REQ_RESERVED;
1068
1069         shared = blk_mq_tag_busy(data.hctx);
1070         rq->tag = blk_mq_get_tag(&data);
1071         if (rq->tag >= 0) {
1072                 if (shared) {
1073                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1074                         atomic_inc(&data.hctx->nr_active);
1075                 }
1076                 data.hctx->tags->rqs[rq->tag] = rq;
1077         }
1078
1079 done:
1080         return rq->tag != -1;
1081 }
1082
1083 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1084                                 int flags, void *key)
1085 {
1086         struct blk_mq_hw_ctx *hctx;
1087
1088         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1089
1090         spin_lock(&hctx->dispatch_wait_lock);
1091         if (!list_empty(&wait->entry)) {
1092                 struct sbitmap_queue *sbq;
1093
1094                 list_del_init(&wait->entry);
1095                 sbq = &hctx->tags->bitmap_tags;
1096                 atomic_dec(&sbq->ws_active);
1097         }
1098         spin_unlock(&hctx->dispatch_wait_lock);
1099
1100         blk_mq_run_hw_queue(hctx, true);
1101         return 1;
1102 }
1103
1104 /*
1105  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1106  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1107  * restart. For both cases, take care to check the condition again after
1108  * marking us as waiting.
1109  */
1110 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1111                                  struct request *rq)
1112 {
1113         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1114         struct wait_queue_head *wq;
1115         wait_queue_entry_t *wait;
1116         bool ret;
1117
1118         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1119                 blk_mq_sched_mark_restart_hctx(hctx);
1120
1121                 /*
1122                  * It's possible that a tag was freed in the window between the
1123                  * allocation failure and adding the hardware queue to the wait
1124                  * queue.
1125                  *
1126                  * Don't clear RESTART here, someone else could have set it.
1127                  * At most this will cost an extra queue run.
1128                  */
1129                 return blk_mq_get_driver_tag(rq);
1130         }
1131
1132         wait = &hctx->dispatch_wait;
1133         if (!list_empty_careful(&wait->entry))
1134                 return false;
1135
1136         wq = &bt_wait_ptr(sbq, hctx)->wait;
1137
1138         spin_lock_irq(&wq->lock);
1139         spin_lock(&hctx->dispatch_wait_lock);
1140         if (!list_empty(&wait->entry)) {
1141                 spin_unlock(&hctx->dispatch_wait_lock);
1142                 spin_unlock_irq(&wq->lock);
1143                 return false;
1144         }
1145
1146         atomic_inc(&sbq->ws_active);
1147         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1148         __add_wait_queue(wq, wait);
1149
1150         /*
1151          * It's possible that a tag was freed in the window between the
1152          * allocation failure and adding the hardware queue to the wait
1153          * queue.
1154          */
1155         ret = blk_mq_get_driver_tag(rq);
1156         if (!ret) {
1157                 spin_unlock(&hctx->dispatch_wait_lock);
1158                 spin_unlock_irq(&wq->lock);
1159                 return false;
1160         }
1161
1162         /*
1163          * We got a tag, remove ourselves from the wait queue to ensure
1164          * someone else gets the wakeup.
1165          */
1166         list_del_init(&wait->entry);
1167         atomic_dec(&sbq->ws_active);
1168         spin_unlock(&hctx->dispatch_wait_lock);
1169         spin_unlock_irq(&wq->lock);
1170
1171         return true;
1172 }
1173
1174 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1175 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1176 /*
1177  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1178  * - EWMA is one simple way to compute running average value
1179  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1180  * - take 4 as factor for avoiding to get too small(0) result, and this
1181  *   factor doesn't matter because EWMA decreases exponentially
1182  */
1183 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1184 {
1185         unsigned int ewma;
1186
1187         if (hctx->queue->elevator)
1188                 return;
1189
1190         ewma = hctx->dispatch_busy;
1191
1192         if (!ewma && !busy)
1193                 return;
1194
1195         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1196         if (busy)
1197                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1198         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1199
1200         hctx->dispatch_busy = ewma;
1201 }
1202
1203 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1204
1205 /*
1206  * Returns true if we did some work AND can potentially do more.
1207  */
1208 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1209                              bool got_budget)
1210 {
1211         struct blk_mq_hw_ctx *hctx;
1212         struct request *rq, *nxt;
1213         bool no_tag = false;
1214         int errors, queued;
1215         blk_status_t ret = BLK_STS_OK;
1216
1217         if (list_empty(list))
1218                 return false;
1219
1220         WARN_ON(!list_is_singular(list) && got_budget);
1221
1222         /*
1223          * Now process all the entries, sending them to the driver.
1224          */
1225         errors = queued = 0;
1226         do {
1227                 struct blk_mq_queue_data bd;
1228
1229                 rq = list_first_entry(list, struct request, queuelist);
1230
1231                 hctx = rq->mq_hctx;
1232                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1233                         break;
1234
1235                 if (!blk_mq_get_driver_tag(rq)) {
1236                         /*
1237                          * The initial allocation attempt failed, so we need to
1238                          * rerun the hardware queue when a tag is freed. The
1239                          * waitqueue takes care of that. If the queue is run
1240                          * before we add this entry back on the dispatch list,
1241                          * we'll re-run it below.
1242                          */
1243                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1244                                 blk_mq_put_dispatch_budget(hctx);
1245                                 /*
1246                                  * For non-shared tags, the RESTART check
1247                                  * will suffice.
1248                                  */
1249                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1250                                         no_tag = true;
1251                                 break;
1252                         }
1253                 }
1254
1255                 list_del_init(&rq->queuelist);
1256
1257                 bd.rq = rq;
1258
1259                 /*
1260                  * Flag last if we have no more requests, or if we have more
1261                  * but can't assign a driver tag to it.
1262                  */
1263                 if (list_empty(list))
1264                         bd.last = true;
1265                 else {
1266                         nxt = list_first_entry(list, struct request, queuelist);
1267                         bd.last = !blk_mq_get_driver_tag(nxt);
1268                 }
1269
1270                 ret = q->mq_ops->queue_rq(hctx, &bd);
1271                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1272                         /*
1273                          * If an I/O scheduler has been configured and we got a
1274                          * driver tag for the next request already, free it
1275                          * again.
1276                          */
1277                         if (!list_empty(list)) {
1278                                 nxt = list_first_entry(list, struct request, queuelist);
1279                                 blk_mq_put_driver_tag(nxt);
1280                         }
1281                         list_add(&rq->queuelist, list);
1282                         __blk_mq_requeue_request(rq);
1283                         break;
1284                 }
1285
1286                 if (unlikely(ret != BLK_STS_OK)) {
1287                         errors++;
1288                         blk_mq_end_request(rq, BLK_STS_IOERR);
1289                         continue;
1290                 }
1291
1292                 queued++;
1293         } while (!list_empty(list));
1294
1295         hctx->dispatched[queued_to_index(queued)]++;
1296
1297         /*
1298          * Any items that need requeuing? Stuff them into hctx->dispatch,
1299          * that is where we will continue on next queue run.
1300          */
1301         if (!list_empty(list)) {
1302                 bool needs_restart;
1303
1304                 /*
1305                  * If we didn't flush the entire list, we could have told
1306                  * the driver there was more coming, but that turned out to
1307                  * be a lie.
1308                  */
1309                 if (q->mq_ops->commit_rqs)
1310                         q->mq_ops->commit_rqs(hctx);
1311
1312                 spin_lock(&hctx->lock);
1313                 list_splice_init(list, &hctx->dispatch);
1314                 spin_unlock(&hctx->lock);
1315
1316                 /*
1317                  * If SCHED_RESTART was set by the caller of this function and
1318                  * it is no longer set that means that it was cleared by another
1319                  * thread and hence that a queue rerun is needed.
1320                  *
1321                  * If 'no_tag' is set, that means that we failed getting
1322                  * a driver tag with an I/O scheduler attached. If our dispatch
1323                  * waitqueue is no longer active, ensure that we run the queue
1324                  * AFTER adding our entries back to the list.
1325                  *
1326                  * If no I/O scheduler has been configured it is possible that
1327                  * the hardware queue got stopped and restarted before requests
1328                  * were pushed back onto the dispatch list. Rerun the queue to
1329                  * avoid starvation. Notes:
1330                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1331                  *   been stopped before rerunning a queue.
1332                  * - Some but not all block drivers stop a queue before
1333                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1334                  *   and dm-rq.
1335                  *
1336                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1337                  * bit is set, run queue after a delay to avoid IO stalls
1338                  * that could otherwise occur if the queue is idle.
1339                  */
1340                 needs_restart = blk_mq_sched_needs_restart(hctx);
1341                 if (!needs_restart ||
1342                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1343                         blk_mq_run_hw_queue(hctx, true);
1344                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1345                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1346
1347                 blk_mq_update_dispatch_busy(hctx, true);
1348                 return false;
1349         } else
1350                 blk_mq_update_dispatch_busy(hctx, false);
1351
1352         /*
1353          * If the host/device is unable to accept more work, inform the
1354          * caller of that.
1355          */
1356         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1357                 return false;
1358
1359         return (queued + errors) != 0;
1360 }
1361
1362 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1363 {
1364         int srcu_idx;
1365
1366         /*
1367          * We should be running this queue from one of the CPUs that
1368          * are mapped to it.
1369          *
1370          * There are at least two related races now between setting
1371          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1372          * __blk_mq_run_hw_queue():
1373          *
1374          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1375          *   but later it becomes online, then this warning is harmless
1376          *   at all
1377          *
1378          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1379          *   but later it becomes offline, then the warning can't be
1380          *   triggered, and we depend on blk-mq timeout handler to
1381          *   handle dispatched requests to this hctx
1382          */
1383         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1384                 cpu_online(hctx->next_cpu)) {
1385                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1386                         raw_smp_processor_id(),
1387                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1388                 dump_stack();
1389         }
1390
1391         /*
1392          * We can't run the queue inline with ints disabled. Ensure that
1393          * we catch bad users of this early.
1394          */
1395         WARN_ON_ONCE(in_interrupt());
1396
1397         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1398
1399         hctx_lock(hctx, &srcu_idx);
1400         blk_mq_sched_dispatch_requests(hctx);
1401         hctx_unlock(hctx, srcu_idx);
1402 }
1403
1404 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1405 {
1406         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1407
1408         if (cpu >= nr_cpu_ids)
1409                 cpu = cpumask_first(hctx->cpumask);
1410         return cpu;
1411 }
1412
1413 /*
1414  * It'd be great if the workqueue API had a way to pass
1415  * in a mask and had some smarts for more clever placement.
1416  * For now we just round-robin here, switching for every
1417  * BLK_MQ_CPU_WORK_BATCH queued items.
1418  */
1419 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1420 {
1421         bool tried = false;
1422         int next_cpu = hctx->next_cpu;
1423
1424         if (hctx->queue->nr_hw_queues == 1)
1425                 return WORK_CPU_UNBOUND;
1426
1427         if (--hctx->next_cpu_batch <= 0) {
1428 select_cpu:
1429                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1430                                 cpu_online_mask);
1431                 if (next_cpu >= nr_cpu_ids)
1432                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1433                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1434         }
1435
1436         /*
1437          * Do unbound schedule if we can't find a online CPU for this hctx,
1438          * and it should only happen in the path of handling CPU DEAD.
1439          */
1440         if (!cpu_online(next_cpu)) {
1441                 if (!tried) {
1442                         tried = true;
1443                         goto select_cpu;
1444                 }
1445
1446                 /*
1447                  * Make sure to re-select CPU next time once after CPUs
1448                  * in hctx->cpumask become online again.
1449                  */
1450                 hctx->next_cpu = next_cpu;
1451                 hctx->next_cpu_batch = 1;
1452                 return WORK_CPU_UNBOUND;
1453         }
1454
1455         hctx->next_cpu = next_cpu;
1456         return next_cpu;
1457 }
1458
1459 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1460                                         unsigned long msecs)
1461 {
1462         if (unlikely(blk_mq_hctx_stopped(hctx)))
1463                 return;
1464
1465         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1466                 int cpu = get_cpu();
1467                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1468                         __blk_mq_run_hw_queue(hctx);
1469                         put_cpu();
1470                         return;
1471                 }
1472
1473                 put_cpu();
1474         }
1475
1476         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1477                                     msecs_to_jiffies(msecs));
1478 }
1479
1480 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1481 {
1482         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1483 }
1484 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1485
1486 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1487 {
1488         int srcu_idx;
1489         bool need_run;
1490
1491         /*
1492          * When queue is quiesced, we may be switching io scheduler, or
1493          * updating nr_hw_queues, or other things, and we can't run queue
1494          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1495          *
1496          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1497          * quiesced.
1498          */
1499         hctx_lock(hctx, &srcu_idx);
1500         need_run = !blk_queue_quiesced(hctx->queue) &&
1501                 blk_mq_hctx_has_pending(hctx);
1502         hctx_unlock(hctx, srcu_idx);
1503
1504         if (need_run) {
1505                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1506                 return true;
1507         }
1508
1509         return false;
1510 }
1511 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1512
1513 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1514 {
1515         struct blk_mq_hw_ctx *hctx;
1516         int i;
1517
1518         queue_for_each_hw_ctx(q, hctx, i) {
1519                 if (blk_mq_hctx_stopped(hctx))
1520                         continue;
1521
1522                 blk_mq_run_hw_queue(hctx, async);
1523         }
1524 }
1525 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1526
1527 /**
1528  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1529  * @q: request queue.
1530  *
1531  * The caller is responsible for serializing this function against
1532  * blk_mq_{start,stop}_hw_queue().
1533  */
1534 bool blk_mq_queue_stopped(struct request_queue *q)
1535 {
1536         struct blk_mq_hw_ctx *hctx;
1537         int i;
1538
1539         queue_for_each_hw_ctx(q, hctx, i)
1540                 if (blk_mq_hctx_stopped(hctx))
1541                         return true;
1542
1543         return false;
1544 }
1545 EXPORT_SYMBOL(blk_mq_queue_stopped);
1546
1547 /*
1548  * This function is often used for pausing .queue_rq() by driver when
1549  * there isn't enough resource or some conditions aren't satisfied, and
1550  * BLK_STS_RESOURCE is usually returned.
1551  *
1552  * We do not guarantee that dispatch can be drained or blocked
1553  * after blk_mq_stop_hw_queue() returns. Please use
1554  * blk_mq_quiesce_queue() for that requirement.
1555  */
1556 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1557 {
1558         cancel_delayed_work(&hctx->run_work);
1559
1560         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1561 }
1562 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1563
1564 /*
1565  * This function is often used for pausing .queue_rq() by driver when
1566  * there isn't enough resource or some conditions aren't satisfied, and
1567  * BLK_STS_RESOURCE is usually returned.
1568  *
1569  * We do not guarantee that dispatch can be drained or blocked
1570  * after blk_mq_stop_hw_queues() returns. Please use
1571  * blk_mq_quiesce_queue() for that requirement.
1572  */
1573 void blk_mq_stop_hw_queues(struct request_queue *q)
1574 {
1575         struct blk_mq_hw_ctx *hctx;
1576         int i;
1577
1578         queue_for_each_hw_ctx(q, hctx, i)
1579                 blk_mq_stop_hw_queue(hctx);
1580 }
1581 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1582
1583 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1584 {
1585         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1586
1587         blk_mq_run_hw_queue(hctx, false);
1588 }
1589 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1590
1591 void blk_mq_start_hw_queues(struct request_queue *q)
1592 {
1593         struct blk_mq_hw_ctx *hctx;
1594         int i;
1595
1596         queue_for_each_hw_ctx(q, hctx, i)
1597                 blk_mq_start_hw_queue(hctx);
1598 }
1599 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1600
1601 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1602 {
1603         if (!blk_mq_hctx_stopped(hctx))
1604                 return;
1605
1606         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1607         blk_mq_run_hw_queue(hctx, async);
1608 }
1609 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1610
1611 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1612 {
1613         struct blk_mq_hw_ctx *hctx;
1614         int i;
1615
1616         queue_for_each_hw_ctx(q, hctx, i)
1617                 blk_mq_start_stopped_hw_queue(hctx, async);
1618 }
1619 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1620
1621 static void blk_mq_run_work_fn(struct work_struct *work)
1622 {
1623         struct blk_mq_hw_ctx *hctx;
1624
1625         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1626
1627         /*
1628          * If we are stopped, don't run the queue.
1629          */
1630         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1631                 return;
1632
1633         __blk_mq_run_hw_queue(hctx);
1634 }
1635
1636 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1637                                             struct request *rq,
1638                                             bool at_head)
1639 {
1640         struct blk_mq_ctx *ctx = rq->mq_ctx;
1641         enum hctx_type type = hctx->type;
1642
1643         lockdep_assert_held(&ctx->lock);
1644
1645         trace_block_rq_insert(hctx->queue, rq);
1646
1647         if (at_head)
1648                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1649         else
1650                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1651 }
1652
1653 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1654                              bool at_head)
1655 {
1656         struct blk_mq_ctx *ctx = rq->mq_ctx;
1657
1658         lockdep_assert_held(&ctx->lock);
1659
1660         __blk_mq_insert_req_list(hctx, rq, at_head);
1661         blk_mq_hctx_mark_pending(hctx, ctx);
1662 }
1663
1664 /*
1665  * Should only be used carefully, when the caller knows we want to
1666  * bypass a potential IO scheduler on the target device.
1667  */
1668 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1669 {
1670         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1671
1672         spin_lock(&hctx->lock);
1673         list_add_tail(&rq->queuelist, &hctx->dispatch);
1674         spin_unlock(&hctx->lock);
1675
1676         if (run_queue)
1677                 blk_mq_run_hw_queue(hctx, false);
1678 }
1679
1680 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1681                             struct list_head *list)
1682
1683 {
1684         struct request *rq;
1685         enum hctx_type type = hctx->type;
1686
1687         /*
1688          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1689          * offline now
1690          */
1691         list_for_each_entry(rq, list, queuelist) {
1692                 BUG_ON(rq->mq_ctx != ctx);
1693                 trace_block_rq_insert(hctx->queue, rq);
1694         }
1695
1696         spin_lock(&ctx->lock);
1697         list_splice_tail_init(list, &ctx->rq_lists[type]);
1698         blk_mq_hctx_mark_pending(hctx, ctx);
1699         spin_unlock(&ctx->lock);
1700 }
1701
1702 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1703 {
1704         struct request *rqa = container_of(a, struct request, queuelist);
1705         struct request *rqb = container_of(b, struct request, queuelist);
1706
1707         if (rqa->mq_ctx < rqb->mq_ctx)
1708                 return -1;
1709         else if (rqa->mq_ctx > rqb->mq_ctx)
1710                 return 1;
1711         else if (rqa->mq_hctx < rqb->mq_hctx)
1712                 return -1;
1713         else if (rqa->mq_hctx > rqb->mq_hctx)
1714                 return 1;
1715
1716         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1717 }
1718
1719 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1720 {
1721         struct blk_mq_hw_ctx *this_hctx;
1722         struct blk_mq_ctx *this_ctx;
1723         struct request_queue *this_q;
1724         struct request *rq;
1725         LIST_HEAD(list);
1726         LIST_HEAD(rq_list);
1727         unsigned int depth;
1728
1729         list_splice_init(&plug->mq_list, &list);
1730
1731         if (plug->rq_count > 2 && plug->multiple_queues)
1732                 list_sort(NULL, &list, plug_rq_cmp);
1733
1734         plug->rq_count = 0;
1735
1736         this_q = NULL;
1737         this_hctx = NULL;
1738         this_ctx = NULL;
1739         depth = 0;
1740
1741         while (!list_empty(&list)) {
1742                 rq = list_entry_rq(list.next);
1743                 list_del_init(&rq->queuelist);
1744                 BUG_ON(!rq->q);
1745                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1746                         if (this_hctx) {
1747                                 trace_block_unplug(this_q, depth, !from_schedule);
1748                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1749                                                                 &rq_list,
1750                                                                 from_schedule);
1751                         }
1752
1753                         this_q = rq->q;
1754                         this_ctx = rq->mq_ctx;
1755                         this_hctx = rq->mq_hctx;
1756                         depth = 0;
1757                 }
1758
1759                 depth++;
1760                 list_add_tail(&rq->queuelist, &rq_list);
1761         }
1762
1763         /*
1764          * If 'this_hctx' is set, we know we have entries to complete
1765          * on 'rq_list'. Do those.
1766          */
1767         if (this_hctx) {
1768                 trace_block_unplug(this_q, depth, !from_schedule);
1769                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1770                                                 from_schedule);
1771         }
1772 }
1773
1774 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1775                 unsigned int nr_segs)
1776 {
1777         if (bio->bi_opf & REQ_RAHEAD)
1778                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1779
1780         rq->__sector = bio->bi_iter.bi_sector;
1781         rq->write_hint = bio->bi_write_hint;
1782         blk_rq_bio_prep(rq, bio, nr_segs);
1783
1784         blk_account_io_start(rq, true);
1785 }
1786
1787 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1788                                             struct request *rq,
1789                                             blk_qc_t *cookie, bool last)
1790 {
1791         struct request_queue *q = rq->q;
1792         struct blk_mq_queue_data bd = {
1793                 .rq = rq,
1794                 .last = last,
1795         };
1796         blk_qc_t new_cookie;
1797         blk_status_t ret;
1798
1799         new_cookie = request_to_qc_t(hctx, rq);
1800
1801         /*
1802          * For OK queue, we are done. For error, caller may kill it.
1803          * Any other error (busy), just add it to our list as we
1804          * previously would have done.
1805          */
1806         ret = q->mq_ops->queue_rq(hctx, &bd);
1807         switch (ret) {
1808         case BLK_STS_OK:
1809                 blk_mq_update_dispatch_busy(hctx, false);
1810                 *cookie = new_cookie;
1811                 break;
1812         case BLK_STS_RESOURCE:
1813         case BLK_STS_DEV_RESOURCE:
1814                 blk_mq_update_dispatch_busy(hctx, true);
1815                 __blk_mq_requeue_request(rq);
1816                 break;
1817         default:
1818                 blk_mq_update_dispatch_busy(hctx, false);
1819                 *cookie = BLK_QC_T_NONE;
1820                 break;
1821         }
1822
1823         return ret;
1824 }
1825
1826 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1827                                                 struct request *rq,
1828                                                 blk_qc_t *cookie,
1829                                                 bool bypass_insert, bool last)
1830 {
1831         struct request_queue *q = rq->q;
1832         bool run_queue = true;
1833
1834         /*
1835          * RCU or SRCU read lock is needed before checking quiesced flag.
1836          *
1837          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1838          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1839          * and avoid driver to try to dispatch again.
1840          */
1841         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1842                 run_queue = false;
1843                 bypass_insert = false;
1844                 goto insert;
1845         }
1846
1847         if (q->elevator && !bypass_insert)
1848                 goto insert;
1849
1850         if (!blk_mq_get_dispatch_budget(hctx))
1851                 goto insert;
1852
1853         if (!blk_mq_get_driver_tag(rq)) {
1854                 blk_mq_put_dispatch_budget(hctx);
1855                 goto insert;
1856         }
1857
1858         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1859 insert:
1860         if (bypass_insert)
1861                 return BLK_STS_RESOURCE;
1862
1863         blk_mq_request_bypass_insert(rq, run_queue);
1864         return BLK_STS_OK;
1865 }
1866
1867 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1868                 struct request *rq, blk_qc_t *cookie)
1869 {
1870         blk_status_t ret;
1871         int srcu_idx;
1872
1873         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1874
1875         hctx_lock(hctx, &srcu_idx);
1876
1877         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1878         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1879                 blk_mq_request_bypass_insert(rq, true);
1880         else if (ret != BLK_STS_OK)
1881                 blk_mq_end_request(rq, ret);
1882
1883         hctx_unlock(hctx, srcu_idx);
1884 }
1885
1886 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1887 {
1888         blk_status_t ret;
1889         int srcu_idx;
1890         blk_qc_t unused_cookie;
1891         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1892
1893         hctx_lock(hctx, &srcu_idx);
1894         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1895         hctx_unlock(hctx, srcu_idx);
1896
1897         return ret;
1898 }
1899
1900 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1901                 struct list_head *list)
1902 {
1903         while (!list_empty(list)) {
1904                 blk_status_t ret;
1905                 struct request *rq = list_first_entry(list, struct request,
1906                                 queuelist);
1907
1908                 list_del_init(&rq->queuelist);
1909                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1910                 if (ret != BLK_STS_OK) {
1911                         if (ret == BLK_STS_RESOURCE ||
1912                                         ret == BLK_STS_DEV_RESOURCE) {
1913                                 blk_mq_request_bypass_insert(rq,
1914                                                         list_empty(list));
1915                                 break;
1916                         }
1917                         blk_mq_end_request(rq, ret);
1918                 }
1919         }
1920
1921         /*
1922          * If we didn't flush the entire list, we could have told
1923          * the driver there was more coming, but that turned out to
1924          * be a lie.
1925          */
1926         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1927                 hctx->queue->mq_ops->commit_rqs(hctx);
1928 }
1929
1930 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1931 {
1932         list_add_tail(&rq->queuelist, &plug->mq_list);
1933         plug->rq_count++;
1934         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1935                 struct request *tmp;
1936
1937                 tmp = list_first_entry(&plug->mq_list, struct request,
1938                                                 queuelist);
1939                 if (tmp->q != rq->q)
1940                         plug->multiple_queues = true;
1941         }
1942 }
1943
1944 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1945 {
1946         const int is_sync = op_is_sync(bio->bi_opf);
1947         const int is_flush_fua = op_is_flush(bio->bi_opf);
1948         struct blk_mq_alloc_data data = { .flags = 0};
1949         struct request *rq;
1950         struct blk_plug *plug;
1951         struct request *same_queue_rq = NULL;
1952         unsigned int nr_segs;
1953         blk_qc_t cookie;
1954
1955         blk_queue_bounce(q, &bio);
1956         __blk_queue_split(q, &bio, &nr_segs);
1957
1958         if (!bio_integrity_prep(bio))
1959                 return BLK_QC_T_NONE;
1960
1961         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1962             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1963                 return BLK_QC_T_NONE;
1964
1965         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1966                 return BLK_QC_T_NONE;
1967
1968         rq_qos_throttle(q, bio);
1969
1970         data.cmd_flags = bio->bi_opf;
1971         rq = blk_mq_get_request(q, bio, &data);
1972         if (unlikely(!rq)) {
1973                 rq_qos_cleanup(q, bio);
1974                 if (bio->bi_opf & REQ_NOWAIT)
1975                         bio_wouldblock_error(bio);
1976                 return BLK_QC_T_NONE;
1977         }
1978
1979         trace_block_getrq(q, bio, bio->bi_opf);
1980
1981         rq_qos_track(q, rq, bio);
1982
1983         cookie = request_to_qc_t(data.hctx, rq);
1984
1985         blk_mq_bio_to_request(rq, bio, nr_segs);
1986
1987         plug = blk_mq_plug(q, bio);
1988         if (unlikely(is_flush_fua)) {
1989                 /* bypass scheduler for flush rq */
1990                 blk_insert_flush(rq);
1991                 blk_mq_run_hw_queue(data.hctx, true);
1992         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1993                 /*
1994                  * Use plugging if we have a ->commit_rqs() hook as well, as
1995                  * we know the driver uses bd->last in a smart fashion.
1996                  */
1997                 unsigned int request_count = plug->rq_count;
1998                 struct request *last = NULL;
1999
2000                 if (!request_count)
2001                         trace_block_plug(q);
2002                 else
2003                         last = list_entry_rq(plug->mq_list.prev);
2004
2005                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2006                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2007                         blk_flush_plug_list(plug, false);
2008                         trace_block_plug(q);
2009                 }
2010
2011                 blk_add_rq_to_plug(plug, rq);
2012         } else if (plug && !blk_queue_nomerges(q)) {
2013                 /*
2014                  * We do limited plugging. If the bio can be merged, do that.
2015                  * Otherwise the existing request in the plug list will be
2016                  * issued. So the plug list will have one request at most
2017                  * The plug list might get flushed before this. If that happens,
2018                  * the plug list is empty, and same_queue_rq is invalid.
2019                  */
2020                 if (list_empty(&plug->mq_list))
2021                         same_queue_rq = NULL;
2022                 if (same_queue_rq) {
2023                         list_del_init(&same_queue_rq->queuelist);
2024                         plug->rq_count--;
2025                 }
2026                 blk_add_rq_to_plug(plug, rq);
2027                 trace_block_plug(q);
2028
2029                 if (same_queue_rq) {
2030                         data.hctx = same_queue_rq->mq_hctx;
2031                         trace_block_unplug(q, 1, true);
2032                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2033                                         &cookie);
2034                 }
2035         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2036                         !data.hctx->dispatch_busy)) {
2037                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2038         } else {
2039                 blk_mq_sched_insert_request(rq, false, true, true);
2040         }
2041
2042         return cookie;
2043 }
2044
2045 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2046                      unsigned int hctx_idx)
2047 {
2048         struct page *page;
2049
2050         if (tags->rqs && set->ops->exit_request) {
2051                 int i;
2052
2053                 for (i = 0; i < tags->nr_tags; i++) {
2054                         struct request *rq = tags->static_rqs[i];
2055
2056                         if (!rq)
2057                                 continue;
2058                         set->ops->exit_request(set, rq, hctx_idx);
2059                         tags->static_rqs[i] = NULL;
2060                 }
2061         }
2062
2063         while (!list_empty(&tags->page_list)) {
2064                 page = list_first_entry(&tags->page_list, struct page, lru);
2065                 list_del_init(&page->lru);
2066                 /*
2067                  * Remove kmemleak object previously allocated in
2068                  * blk_mq_alloc_rqs().
2069                  */
2070                 kmemleak_free(page_address(page));
2071                 __free_pages(page, page->private);
2072         }
2073 }
2074
2075 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2076 {
2077         kfree(tags->rqs);
2078         tags->rqs = NULL;
2079         kfree(tags->static_rqs);
2080         tags->static_rqs = NULL;
2081
2082         blk_mq_free_tags(tags);
2083 }
2084
2085 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2086                                         unsigned int hctx_idx,
2087                                         unsigned int nr_tags,
2088                                         unsigned int reserved_tags)
2089 {
2090         struct blk_mq_tags *tags;
2091         int node;
2092
2093         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2094         if (node == NUMA_NO_NODE)
2095                 node = set->numa_node;
2096
2097         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2098                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2099         if (!tags)
2100                 return NULL;
2101
2102         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2103                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2104                                  node);
2105         if (!tags->rqs) {
2106                 blk_mq_free_tags(tags);
2107                 return NULL;
2108         }
2109
2110         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2111                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2112                                         node);
2113         if (!tags->static_rqs) {
2114                 kfree(tags->rqs);
2115                 blk_mq_free_tags(tags);
2116                 return NULL;
2117         }
2118
2119         return tags;
2120 }
2121
2122 static size_t order_to_size(unsigned int order)
2123 {
2124         return (size_t)PAGE_SIZE << order;
2125 }
2126
2127 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2128                                unsigned int hctx_idx, int node)
2129 {
2130         int ret;
2131
2132         if (set->ops->init_request) {
2133                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2134                 if (ret)
2135                         return ret;
2136         }
2137
2138         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2139         return 0;
2140 }
2141
2142 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2143                      unsigned int hctx_idx, unsigned int depth)
2144 {
2145         unsigned int i, j, entries_per_page, max_order = 4;
2146         size_t rq_size, left;
2147         int node;
2148
2149         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2150         if (node == NUMA_NO_NODE)
2151                 node = set->numa_node;
2152
2153         INIT_LIST_HEAD(&tags->page_list);
2154
2155         /*
2156          * rq_size is the size of the request plus driver payload, rounded
2157          * to the cacheline size
2158          */
2159         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2160                                 cache_line_size());
2161         left = rq_size * depth;
2162
2163         for (i = 0; i < depth; ) {
2164                 int this_order = max_order;
2165                 struct page *page;
2166                 int to_do;
2167                 void *p;
2168
2169                 while (this_order && left < order_to_size(this_order - 1))
2170                         this_order--;
2171
2172                 do {
2173                         page = alloc_pages_node(node,
2174                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2175                                 this_order);
2176                         if (page)
2177                                 break;
2178                         if (!this_order--)
2179                                 break;
2180                         if (order_to_size(this_order) < rq_size)
2181                                 break;
2182                 } while (1);
2183
2184                 if (!page)
2185                         goto fail;
2186
2187                 page->private = this_order;
2188                 list_add_tail(&page->lru, &tags->page_list);
2189
2190                 p = page_address(page);
2191                 /*
2192                  * Allow kmemleak to scan these pages as they contain pointers
2193                  * to additional allocations like via ops->init_request().
2194                  */
2195                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2196                 entries_per_page = order_to_size(this_order) / rq_size;
2197                 to_do = min(entries_per_page, depth - i);
2198                 left -= to_do * rq_size;
2199                 for (j = 0; j < to_do; j++) {
2200                         struct request *rq = p;
2201
2202                         tags->static_rqs[i] = rq;
2203                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2204                                 tags->static_rqs[i] = NULL;
2205                                 goto fail;
2206                         }
2207
2208                         p += rq_size;
2209                         i++;
2210                 }
2211         }
2212         return 0;
2213
2214 fail:
2215         blk_mq_free_rqs(set, tags, hctx_idx);
2216         return -ENOMEM;
2217 }
2218
2219 /*
2220  * 'cpu' is going away. splice any existing rq_list entries from this
2221  * software queue to the hw queue dispatch list, and ensure that it
2222  * gets run.
2223  */
2224 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2225 {
2226         struct blk_mq_hw_ctx *hctx;
2227         struct blk_mq_ctx *ctx;
2228         LIST_HEAD(tmp);
2229         enum hctx_type type;
2230
2231         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2232         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2233         type = hctx->type;
2234
2235         spin_lock(&ctx->lock);
2236         if (!list_empty(&ctx->rq_lists[type])) {
2237                 list_splice_init(&ctx->rq_lists[type], &tmp);
2238                 blk_mq_hctx_clear_pending(hctx, ctx);
2239         }
2240         spin_unlock(&ctx->lock);
2241
2242         if (list_empty(&tmp))
2243                 return 0;
2244
2245         spin_lock(&hctx->lock);
2246         list_splice_tail_init(&tmp, &hctx->dispatch);
2247         spin_unlock(&hctx->lock);
2248
2249         blk_mq_run_hw_queue(hctx, true);
2250         return 0;
2251 }
2252
2253 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2254 {
2255         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2256                                             &hctx->cpuhp_dead);
2257 }
2258
2259 /* hctx->ctxs will be freed in queue's release handler */
2260 static void blk_mq_exit_hctx(struct request_queue *q,
2261                 struct blk_mq_tag_set *set,
2262                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2263 {
2264         if (blk_mq_hw_queue_mapped(hctx))
2265                 blk_mq_tag_idle(hctx);
2266
2267         if (set->ops->exit_request)
2268                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2269
2270         if (set->ops->exit_hctx)
2271                 set->ops->exit_hctx(hctx, hctx_idx);
2272
2273         blk_mq_remove_cpuhp(hctx);
2274
2275         spin_lock(&q->unused_hctx_lock);
2276         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2277         spin_unlock(&q->unused_hctx_lock);
2278 }
2279
2280 static void blk_mq_exit_hw_queues(struct request_queue *q,
2281                 struct blk_mq_tag_set *set, int nr_queue)
2282 {
2283         struct blk_mq_hw_ctx *hctx;
2284         unsigned int i;
2285
2286         queue_for_each_hw_ctx(q, hctx, i) {
2287                 if (i == nr_queue)
2288                         break;
2289                 blk_mq_debugfs_unregister_hctx(hctx);
2290                 blk_mq_exit_hctx(q, set, hctx, i);
2291         }
2292 }
2293
2294 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2295 {
2296         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2297
2298         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2299                            __alignof__(struct blk_mq_hw_ctx)) !=
2300                      sizeof(struct blk_mq_hw_ctx));
2301
2302         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2303                 hw_ctx_size += sizeof(struct srcu_struct);
2304
2305         return hw_ctx_size;
2306 }
2307
2308 static int blk_mq_init_hctx(struct request_queue *q,
2309                 struct blk_mq_tag_set *set,
2310                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2311 {
2312         hctx->queue_num = hctx_idx;
2313
2314         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2315
2316         hctx->tags = set->tags[hctx_idx];
2317
2318         if (set->ops->init_hctx &&
2319             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2320                 goto unregister_cpu_notifier;
2321
2322         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2323                                 hctx->numa_node))
2324                 goto exit_hctx;
2325         return 0;
2326
2327  exit_hctx:
2328         if (set->ops->exit_hctx)
2329                 set->ops->exit_hctx(hctx, hctx_idx);
2330  unregister_cpu_notifier:
2331         blk_mq_remove_cpuhp(hctx);
2332         return -1;
2333 }
2334
2335 static struct blk_mq_hw_ctx *
2336 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2337                 int node)
2338 {
2339         struct blk_mq_hw_ctx *hctx;
2340         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2341
2342         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2343         if (!hctx)
2344                 goto fail_alloc_hctx;
2345
2346         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2347                 goto free_hctx;
2348
2349         atomic_set(&hctx->nr_active, 0);
2350         if (node == NUMA_NO_NODE)
2351                 node = set->numa_node;
2352         hctx->numa_node = node;
2353
2354         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2355         spin_lock_init(&hctx->lock);
2356         INIT_LIST_HEAD(&hctx->dispatch);
2357         hctx->queue = q;
2358         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2359
2360         INIT_LIST_HEAD(&hctx->hctx_list);
2361
2362         /*
2363          * Allocate space for all possible cpus to avoid allocation at
2364          * runtime
2365          */
2366         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2367                         gfp, node);
2368         if (!hctx->ctxs)
2369                 goto free_cpumask;
2370
2371         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2372                                 gfp, node))
2373                 goto free_ctxs;
2374         hctx->nr_ctx = 0;
2375
2376         spin_lock_init(&hctx->dispatch_wait_lock);
2377         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2378         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2379
2380         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2381                         gfp);
2382         if (!hctx->fq)
2383                 goto free_bitmap;
2384
2385         if (hctx->flags & BLK_MQ_F_BLOCKING)
2386                 init_srcu_struct(hctx->srcu);
2387         blk_mq_hctx_kobj_init(hctx);
2388
2389         return hctx;
2390
2391  free_bitmap:
2392         sbitmap_free(&hctx->ctx_map);
2393  free_ctxs:
2394         kfree(hctx->ctxs);
2395  free_cpumask:
2396         free_cpumask_var(hctx->cpumask);
2397  free_hctx:
2398         kfree(hctx);
2399  fail_alloc_hctx:
2400         return NULL;
2401 }
2402
2403 static void blk_mq_init_cpu_queues(struct request_queue *q,
2404                                    unsigned int nr_hw_queues)
2405 {
2406         struct blk_mq_tag_set *set = q->tag_set;
2407         unsigned int i, j;
2408
2409         for_each_possible_cpu(i) {
2410                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2411                 struct blk_mq_hw_ctx *hctx;
2412                 int k;
2413
2414                 __ctx->cpu = i;
2415                 spin_lock_init(&__ctx->lock);
2416                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2417                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2418
2419                 __ctx->queue = q;
2420
2421                 /*
2422                  * Set local node, IFF we have more than one hw queue. If
2423                  * not, we remain on the home node of the device
2424                  */
2425                 for (j = 0; j < set->nr_maps; j++) {
2426                         hctx = blk_mq_map_queue_type(q, j, i);
2427                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2428                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2429                 }
2430         }
2431 }
2432
2433 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2434 {
2435         int ret = 0;
2436
2437         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2438                                         set->queue_depth, set->reserved_tags);
2439         if (!set->tags[hctx_idx])
2440                 return false;
2441
2442         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2443                                 set->queue_depth);
2444         if (!ret)
2445                 return true;
2446
2447         blk_mq_free_rq_map(set->tags[hctx_idx]);
2448         set->tags[hctx_idx] = NULL;
2449         return false;
2450 }
2451
2452 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2453                                          unsigned int hctx_idx)
2454 {
2455         if (set->tags && set->tags[hctx_idx]) {
2456                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2457                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2458                 set->tags[hctx_idx] = NULL;
2459         }
2460 }
2461
2462 static void blk_mq_map_swqueue(struct request_queue *q)
2463 {
2464         unsigned int i, j, hctx_idx;
2465         struct blk_mq_hw_ctx *hctx;
2466         struct blk_mq_ctx *ctx;
2467         struct blk_mq_tag_set *set = q->tag_set;
2468
2469         queue_for_each_hw_ctx(q, hctx, i) {
2470                 cpumask_clear(hctx->cpumask);
2471                 hctx->nr_ctx = 0;
2472                 hctx->dispatch_from = NULL;
2473         }
2474
2475         /*
2476          * Map software to hardware queues.
2477          *
2478          * If the cpu isn't present, the cpu is mapped to first hctx.
2479          */
2480         for_each_possible_cpu(i) {
2481                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2482                 /* unmapped hw queue can be remapped after CPU topo changed */
2483                 if (!set->tags[hctx_idx] &&
2484                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2485                         /*
2486                          * If tags initialization fail for some hctx,
2487                          * that hctx won't be brought online.  In this
2488                          * case, remap the current ctx to hctx[0] which
2489                          * is guaranteed to always have tags allocated
2490                          */
2491                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2492                 }
2493
2494                 ctx = per_cpu_ptr(q->queue_ctx, i);
2495                 for (j = 0; j < set->nr_maps; j++) {
2496                         if (!set->map[j].nr_queues) {
2497                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2498                                                 HCTX_TYPE_DEFAULT, i);
2499                                 continue;
2500                         }
2501
2502                         hctx = blk_mq_map_queue_type(q, j, i);
2503                         ctx->hctxs[j] = hctx;
2504                         /*
2505                          * If the CPU is already set in the mask, then we've
2506                          * mapped this one already. This can happen if
2507                          * devices share queues across queue maps.
2508                          */
2509                         if (cpumask_test_cpu(i, hctx->cpumask))
2510                                 continue;
2511
2512                         cpumask_set_cpu(i, hctx->cpumask);
2513                         hctx->type = j;
2514                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2515                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2516
2517                         /*
2518                          * If the nr_ctx type overflows, we have exceeded the
2519                          * amount of sw queues we can support.
2520                          */
2521                         BUG_ON(!hctx->nr_ctx);
2522                 }
2523
2524                 for (; j < HCTX_MAX_TYPES; j++)
2525                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2526                                         HCTX_TYPE_DEFAULT, i);
2527         }
2528
2529         queue_for_each_hw_ctx(q, hctx, i) {
2530                 /*
2531                  * If no software queues are mapped to this hardware queue,
2532                  * disable it and free the request entries.
2533                  */
2534                 if (!hctx->nr_ctx) {
2535                         /* Never unmap queue 0.  We need it as a
2536                          * fallback in case of a new remap fails
2537                          * allocation
2538                          */
2539                         if (i && set->tags[i])
2540                                 blk_mq_free_map_and_requests(set, i);
2541
2542                         hctx->tags = NULL;
2543                         continue;
2544                 }
2545
2546                 hctx->tags = set->tags[i];
2547                 WARN_ON(!hctx->tags);
2548
2549                 /*
2550                  * Set the map size to the number of mapped software queues.
2551                  * This is more accurate and more efficient than looping
2552                  * over all possibly mapped software queues.
2553                  */
2554                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2555
2556                 /*
2557                  * Initialize batch roundrobin counts
2558                  */
2559                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2560                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2561         }
2562 }
2563
2564 /*
2565  * Caller needs to ensure that we're either frozen/quiesced, or that
2566  * the queue isn't live yet.
2567  */
2568 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2569 {
2570         struct blk_mq_hw_ctx *hctx;
2571         int i;
2572
2573         queue_for_each_hw_ctx(q, hctx, i) {
2574                 if (shared)
2575                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2576                 else
2577                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2578         }
2579 }
2580
2581 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2582                                         bool shared)
2583 {
2584         struct request_queue *q;
2585
2586         lockdep_assert_held(&set->tag_list_lock);
2587
2588         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2589                 blk_mq_freeze_queue(q);
2590                 queue_set_hctx_shared(q, shared);
2591                 blk_mq_unfreeze_queue(q);
2592         }
2593 }
2594
2595 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2596 {
2597         struct blk_mq_tag_set *set = q->tag_set;
2598
2599         mutex_lock(&set->tag_list_lock);
2600         list_del_rcu(&q->tag_set_list);
2601         if (list_is_singular(&set->tag_list)) {
2602                 /* just transitioned to unshared */
2603                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2604                 /* update existing queue */
2605                 blk_mq_update_tag_set_depth(set, false);
2606         }
2607         mutex_unlock(&set->tag_list_lock);
2608         INIT_LIST_HEAD(&q->tag_set_list);
2609 }
2610
2611 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2612                                      struct request_queue *q)
2613 {
2614         mutex_lock(&set->tag_list_lock);
2615
2616         /*
2617          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2618          */
2619         if (!list_empty(&set->tag_list) &&
2620             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2621                 set->flags |= BLK_MQ_F_TAG_SHARED;
2622                 /* update existing queue */
2623                 blk_mq_update_tag_set_depth(set, true);
2624         }
2625         if (set->flags & BLK_MQ_F_TAG_SHARED)
2626                 queue_set_hctx_shared(q, true);
2627         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2628
2629         mutex_unlock(&set->tag_list_lock);
2630 }
2631
2632 /* All allocations will be freed in release handler of q->mq_kobj */
2633 static int blk_mq_alloc_ctxs(struct request_queue *q)
2634 {
2635         struct blk_mq_ctxs *ctxs;
2636         int cpu;
2637
2638         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2639         if (!ctxs)
2640                 return -ENOMEM;
2641
2642         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2643         if (!ctxs->queue_ctx)
2644                 goto fail;
2645
2646         for_each_possible_cpu(cpu) {
2647                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2648                 ctx->ctxs = ctxs;
2649         }
2650
2651         q->mq_kobj = &ctxs->kobj;
2652         q->queue_ctx = ctxs->queue_ctx;
2653
2654         return 0;
2655  fail:
2656         kfree(ctxs);
2657         return -ENOMEM;
2658 }
2659
2660 /*
2661  * It is the actual release handler for mq, but we do it from
2662  * request queue's release handler for avoiding use-after-free
2663  * and headache because q->mq_kobj shouldn't have been introduced,
2664  * but we can't group ctx/kctx kobj without it.
2665  */
2666 void blk_mq_release(struct request_queue *q)
2667 {
2668         struct blk_mq_hw_ctx *hctx, *next;
2669         int i;
2670
2671         queue_for_each_hw_ctx(q, hctx, i)
2672                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2673
2674         /* all hctx are in .unused_hctx_list now */
2675         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2676                 list_del_init(&hctx->hctx_list);
2677                 kobject_put(&hctx->kobj);
2678         }
2679
2680         kfree(q->queue_hw_ctx);
2681
2682         /*
2683          * release .mq_kobj and sw queue's kobject now because
2684          * both share lifetime with request queue.
2685          */
2686         blk_mq_sysfs_deinit(q);
2687 }
2688
2689 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2690 {
2691         struct request_queue *uninit_q, *q;
2692
2693         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2694         if (!uninit_q)
2695                 return ERR_PTR(-ENOMEM);
2696
2697         /*
2698          * Initialize the queue without an elevator. device_add_disk() will do
2699          * the initialization.
2700          */
2701         q = blk_mq_init_allocated_queue(set, uninit_q, false);
2702         if (IS_ERR(q))
2703                 blk_cleanup_queue(uninit_q);
2704
2705         return q;
2706 }
2707 EXPORT_SYMBOL(blk_mq_init_queue);
2708
2709 /*
2710  * Helper for setting up a queue with mq ops, given queue depth, and
2711  * the passed in mq ops flags.
2712  */
2713 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2714                                            const struct blk_mq_ops *ops,
2715                                            unsigned int queue_depth,
2716                                            unsigned int set_flags)
2717 {
2718         struct request_queue *q;
2719         int ret;
2720
2721         memset(set, 0, sizeof(*set));
2722         set->ops = ops;
2723         set->nr_hw_queues = 1;
2724         set->nr_maps = 1;
2725         set->queue_depth = queue_depth;
2726         set->numa_node = NUMA_NO_NODE;
2727         set->flags = set_flags;
2728
2729         ret = blk_mq_alloc_tag_set(set);
2730         if (ret)
2731                 return ERR_PTR(ret);
2732
2733         q = blk_mq_init_queue(set);
2734         if (IS_ERR(q)) {
2735                 blk_mq_free_tag_set(set);
2736                 return q;
2737         }
2738
2739         return q;
2740 }
2741 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2742
2743 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2744                 struct blk_mq_tag_set *set, struct request_queue *q,
2745                 int hctx_idx, int node)
2746 {
2747         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2748
2749         /* reuse dead hctx first */
2750         spin_lock(&q->unused_hctx_lock);
2751         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2752                 if (tmp->numa_node == node) {
2753                         hctx = tmp;
2754                         break;
2755                 }
2756         }
2757         if (hctx)
2758                 list_del_init(&hctx->hctx_list);
2759         spin_unlock(&q->unused_hctx_lock);
2760
2761         if (!hctx)
2762                 hctx = blk_mq_alloc_hctx(q, set, node);
2763         if (!hctx)
2764                 goto fail;
2765
2766         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2767                 goto free_hctx;
2768
2769         return hctx;
2770
2771  free_hctx:
2772         kobject_put(&hctx->kobj);
2773  fail:
2774         return NULL;
2775 }
2776
2777 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2778                                                 struct request_queue *q)
2779 {
2780         int i, j, end;
2781         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2782
2783         /* protect against switching io scheduler  */
2784         mutex_lock(&q->sysfs_lock);
2785         for (i = 0; i < set->nr_hw_queues; i++) {
2786                 int node;
2787                 struct blk_mq_hw_ctx *hctx;
2788
2789                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2790                 /*
2791                  * If the hw queue has been mapped to another numa node,
2792                  * we need to realloc the hctx. If allocation fails, fallback
2793                  * to use the previous one.
2794                  */
2795                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2796                         continue;
2797
2798                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2799                 if (hctx) {
2800                         if (hctxs[i])
2801                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2802                         hctxs[i] = hctx;
2803                 } else {
2804                         if (hctxs[i])
2805                                 pr_warn("Allocate new hctx on node %d fails,\
2806                                                 fallback to previous one on node %d\n",
2807                                                 node, hctxs[i]->numa_node);
2808                         else
2809                                 break;
2810                 }
2811         }
2812         /*
2813          * Increasing nr_hw_queues fails. Free the newly allocated
2814          * hctxs and keep the previous q->nr_hw_queues.
2815          */
2816         if (i != set->nr_hw_queues) {
2817                 j = q->nr_hw_queues;
2818                 end = i;
2819         } else {
2820                 j = i;
2821                 end = q->nr_hw_queues;
2822                 q->nr_hw_queues = set->nr_hw_queues;
2823         }
2824
2825         for (; j < end; j++) {
2826                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2827
2828                 if (hctx) {
2829                         if (hctx->tags)
2830                                 blk_mq_free_map_and_requests(set, j);
2831                         blk_mq_exit_hctx(q, set, hctx, j);
2832                         hctxs[j] = NULL;
2833                 }
2834         }
2835         mutex_unlock(&q->sysfs_lock);
2836 }
2837
2838 /*
2839  * Maximum number of hardware queues we support. For single sets, we'll never
2840  * have more than the CPUs (software queues). For multiple sets, the tag_set
2841  * user may have set ->nr_hw_queues larger.
2842  */
2843 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2844 {
2845         if (set->nr_maps == 1)
2846                 return nr_cpu_ids;
2847
2848         return max(set->nr_hw_queues, nr_cpu_ids);
2849 }
2850
2851 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2852                                                   struct request_queue *q,
2853                                                   bool elevator_init)
2854 {
2855         /* mark the queue as mq asap */
2856         q->mq_ops = set->ops;
2857
2858         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2859                                              blk_mq_poll_stats_bkt,
2860                                              BLK_MQ_POLL_STATS_BKTS, q);
2861         if (!q->poll_cb)
2862                 goto err_exit;
2863
2864         if (blk_mq_alloc_ctxs(q))
2865                 goto err_poll;
2866
2867         /* init q->mq_kobj and sw queues' kobjects */
2868         blk_mq_sysfs_init(q);
2869
2870         q->nr_queues = nr_hw_queues(set);
2871         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2872                                                 GFP_KERNEL, set->numa_node);
2873         if (!q->queue_hw_ctx)
2874                 goto err_sys_init;
2875
2876         INIT_LIST_HEAD(&q->unused_hctx_list);
2877         spin_lock_init(&q->unused_hctx_lock);
2878
2879         blk_mq_realloc_hw_ctxs(set, q);
2880         if (!q->nr_hw_queues)
2881                 goto err_hctxs;
2882
2883         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2884         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2885
2886         q->tag_set = set;
2887
2888         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2889         if (set->nr_maps > HCTX_TYPE_POLL &&
2890             set->map[HCTX_TYPE_POLL].nr_queues)
2891                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2892
2893         q->sg_reserved_size = INT_MAX;
2894
2895         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2896         INIT_LIST_HEAD(&q->requeue_list);
2897         spin_lock_init(&q->requeue_lock);
2898
2899         blk_queue_make_request(q, blk_mq_make_request);
2900
2901         /*
2902          * Do this after blk_queue_make_request() overrides it...
2903          */
2904         q->nr_requests = set->queue_depth;
2905
2906         /*
2907          * Default to classic polling
2908          */
2909         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2910
2911         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2912         blk_mq_add_queue_tag_set(set, q);
2913         blk_mq_map_swqueue(q);
2914
2915         if (elevator_init)
2916                 elevator_init_mq(q);
2917
2918         return q;
2919
2920 err_hctxs:
2921         kfree(q->queue_hw_ctx);
2922         q->nr_hw_queues = 0;
2923 err_sys_init:
2924         blk_mq_sysfs_deinit(q);
2925 err_poll:
2926         blk_stat_free_callback(q->poll_cb);
2927         q->poll_cb = NULL;
2928 err_exit:
2929         q->mq_ops = NULL;
2930         return ERR_PTR(-ENOMEM);
2931 }
2932 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2933
2934 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2935 void blk_mq_exit_queue(struct request_queue *q)
2936 {
2937         struct blk_mq_tag_set   *set = q->tag_set;
2938
2939         blk_mq_del_queue_tag_set(q);
2940         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2941 }
2942
2943 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2944 {
2945         int i;
2946
2947         for (i = 0; i < set->nr_hw_queues; i++)
2948                 if (!__blk_mq_alloc_rq_map(set, i))
2949                         goto out_unwind;
2950
2951         return 0;
2952
2953 out_unwind:
2954         while (--i >= 0)
2955                 blk_mq_free_rq_map(set->tags[i]);
2956
2957         return -ENOMEM;
2958 }
2959
2960 /*
2961  * Allocate the request maps associated with this tag_set. Note that this
2962  * may reduce the depth asked for, if memory is tight. set->queue_depth
2963  * will be updated to reflect the allocated depth.
2964  */
2965 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2966 {
2967         unsigned int depth;
2968         int err;
2969
2970         depth = set->queue_depth;
2971         do {
2972                 err = __blk_mq_alloc_rq_maps(set);
2973                 if (!err)
2974                         break;
2975
2976                 set->queue_depth >>= 1;
2977                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2978                         err = -ENOMEM;
2979                         break;
2980                 }
2981         } while (set->queue_depth);
2982
2983         if (!set->queue_depth || err) {
2984                 pr_err("blk-mq: failed to allocate request map\n");
2985                 return -ENOMEM;
2986         }
2987
2988         if (depth != set->queue_depth)
2989                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2990                                                 depth, set->queue_depth);
2991
2992         return 0;
2993 }
2994
2995 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2996 {
2997         if (set->ops->map_queues && !is_kdump_kernel()) {
2998                 int i;
2999
3000                 /*
3001                  * transport .map_queues is usually done in the following
3002                  * way:
3003                  *
3004                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3005                  *      mask = get_cpu_mask(queue)
3006                  *      for_each_cpu(cpu, mask)
3007                  *              set->map[x].mq_map[cpu] = queue;
3008                  * }
3009                  *
3010                  * When we need to remap, the table has to be cleared for
3011                  * killing stale mapping since one CPU may not be mapped
3012                  * to any hw queue.
3013                  */
3014                 for (i = 0; i < set->nr_maps; i++)
3015                         blk_mq_clear_mq_map(&set->map[i]);
3016
3017                 return set->ops->map_queues(set);
3018         } else {
3019                 BUG_ON(set->nr_maps > 1);
3020                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3021         }
3022 }
3023
3024 /*
3025  * Alloc a tag set to be associated with one or more request queues.
3026  * May fail with EINVAL for various error conditions. May adjust the
3027  * requested depth down, if it's too large. In that case, the set
3028  * value will be stored in set->queue_depth.
3029  */
3030 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3031 {
3032         int i, ret;
3033
3034         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3035
3036         if (!set->nr_hw_queues)
3037                 return -EINVAL;
3038         if (!set->queue_depth)
3039                 return -EINVAL;
3040         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3041                 return -EINVAL;
3042
3043         if (!set->ops->queue_rq)
3044                 return -EINVAL;
3045
3046         if (!set->ops->get_budget ^ !set->ops->put_budget)
3047                 return -EINVAL;
3048
3049         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3050                 pr_info("blk-mq: reduced tag depth to %u\n",
3051                         BLK_MQ_MAX_DEPTH);
3052                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3053         }
3054
3055         if (!set->nr_maps)
3056                 set->nr_maps = 1;
3057         else if (set->nr_maps > HCTX_MAX_TYPES)
3058                 return -EINVAL;
3059
3060         /*
3061          * If a crashdump is active, then we are potentially in a very
3062          * memory constrained environment. Limit us to 1 queue and
3063          * 64 tags to prevent using too much memory.
3064          */
3065         if (is_kdump_kernel()) {
3066                 set->nr_hw_queues = 1;
3067                 set->nr_maps = 1;
3068                 set->queue_depth = min(64U, set->queue_depth);
3069         }
3070         /*
3071          * There is no use for more h/w queues than cpus if we just have
3072          * a single map
3073          */
3074         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3075                 set->nr_hw_queues = nr_cpu_ids;
3076
3077         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3078                                  GFP_KERNEL, set->numa_node);
3079         if (!set->tags)
3080                 return -ENOMEM;
3081
3082         ret = -ENOMEM;
3083         for (i = 0; i < set->nr_maps; i++) {
3084                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3085                                                   sizeof(set->map[i].mq_map[0]),
3086                                                   GFP_KERNEL, set->numa_node);
3087                 if (!set->map[i].mq_map)
3088                         goto out_free_mq_map;
3089                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3090         }
3091
3092         ret = blk_mq_update_queue_map(set);
3093         if (ret)
3094                 goto out_free_mq_map;
3095
3096         ret = blk_mq_alloc_rq_maps(set);
3097         if (ret)
3098                 goto out_free_mq_map;
3099
3100         mutex_init(&set->tag_list_lock);
3101         INIT_LIST_HEAD(&set->tag_list);
3102
3103         return 0;
3104
3105 out_free_mq_map:
3106         for (i = 0; i < set->nr_maps; i++) {
3107                 kfree(set->map[i].mq_map);
3108                 set->map[i].mq_map = NULL;
3109         }
3110         kfree(set->tags);
3111         set->tags = NULL;
3112         return ret;
3113 }
3114 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3115
3116 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3117 {
3118         int i, j;
3119
3120         for (i = 0; i < nr_hw_queues(set); i++)
3121                 blk_mq_free_map_and_requests(set, i);
3122
3123         for (j = 0; j < set->nr_maps; j++) {
3124                 kfree(set->map[j].mq_map);
3125                 set->map[j].mq_map = NULL;
3126         }
3127
3128         kfree(set->tags);
3129         set->tags = NULL;
3130 }
3131 EXPORT_SYMBOL(blk_mq_free_tag_set);
3132
3133 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3134 {
3135         struct blk_mq_tag_set *set = q->tag_set;
3136         struct blk_mq_hw_ctx *hctx;
3137         int i, ret;
3138
3139         if (!set)
3140                 return -EINVAL;
3141
3142         if (q->nr_requests == nr)
3143                 return 0;
3144
3145         blk_mq_freeze_queue(q);
3146         blk_mq_quiesce_queue(q);
3147
3148         ret = 0;
3149         queue_for_each_hw_ctx(q, hctx, i) {
3150                 if (!hctx->tags)
3151                         continue;
3152                 /*
3153                  * If we're using an MQ scheduler, just update the scheduler
3154                  * queue depth. This is similar to what the old code would do.
3155                  */
3156                 if (!hctx->sched_tags) {
3157                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3158                                                         false);
3159                 } else {
3160                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3161                                                         nr, true);
3162                 }
3163                 if (ret)
3164                         break;
3165                 if (q->elevator && q->elevator->type->ops.depth_updated)
3166                         q->elevator->type->ops.depth_updated(hctx);
3167         }
3168
3169         if (!ret)
3170                 q->nr_requests = nr;
3171
3172         blk_mq_unquiesce_queue(q);
3173         blk_mq_unfreeze_queue(q);
3174
3175         return ret;
3176 }
3177
3178 /*
3179  * request_queue and elevator_type pair.
3180  * It is just used by __blk_mq_update_nr_hw_queues to cache
3181  * the elevator_type associated with a request_queue.
3182  */
3183 struct blk_mq_qe_pair {
3184         struct list_head node;
3185         struct request_queue *q;
3186         struct elevator_type *type;
3187 };
3188
3189 /*
3190  * Cache the elevator_type in qe pair list and switch the
3191  * io scheduler to 'none'
3192  */
3193 static bool blk_mq_elv_switch_none(struct list_head *head,
3194                 struct request_queue *q)
3195 {
3196         struct blk_mq_qe_pair *qe;
3197
3198         if (!q->elevator)
3199                 return true;
3200
3201         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3202         if (!qe)
3203                 return false;
3204
3205         INIT_LIST_HEAD(&qe->node);
3206         qe->q = q;
3207         qe->type = q->elevator->type;
3208         list_add(&qe->node, head);
3209
3210         mutex_lock(&q->sysfs_lock);
3211         /*
3212          * After elevator_switch_mq, the previous elevator_queue will be
3213          * released by elevator_release. The reference of the io scheduler
3214          * module get by elevator_get will also be put. So we need to get
3215          * a reference of the io scheduler module here to prevent it to be
3216          * removed.
3217          */
3218         __module_get(qe->type->elevator_owner);
3219         elevator_switch_mq(q, NULL);
3220         mutex_unlock(&q->sysfs_lock);
3221
3222         return true;
3223 }
3224
3225 static void blk_mq_elv_switch_back(struct list_head *head,
3226                 struct request_queue *q)
3227 {
3228         struct blk_mq_qe_pair *qe;
3229         struct elevator_type *t = NULL;
3230
3231         list_for_each_entry(qe, head, node)
3232                 if (qe->q == q) {
3233                         t = qe->type;
3234                         break;
3235                 }
3236
3237         if (!t)
3238                 return;
3239
3240         list_del(&qe->node);
3241         kfree(qe);
3242
3243         mutex_lock(&q->sysfs_lock);
3244         elevator_switch_mq(q, t);
3245         mutex_unlock(&q->sysfs_lock);
3246 }
3247
3248 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3249                                                         int nr_hw_queues)
3250 {
3251         struct request_queue *q;
3252         LIST_HEAD(head);
3253         int prev_nr_hw_queues;
3254
3255         lockdep_assert_held(&set->tag_list_lock);
3256
3257         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3258                 nr_hw_queues = nr_cpu_ids;
3259         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3260                 return;
3261
3262         list_for_each_entry(q, &set->tag_list, tag_set_list)
3263                 blk_mq_freeze_queue(q);
3264         /*
3265          * Sync with blk_mq_queue_tag_busy_iter.
3266          */
3267         synchronize_rcu();
3268         /*
3269          * Switch IO scheduler to 'none', cleaning up the data associated
3270          * with the previous scheduler. We will switch back once we are done
3271          * updating the new sw to hw queue mappings.
3272          */
3273         list_for_each_entry(q, &set->tag_list, tag_set_list)
3274                 if (!blk_mq_elv_switch_none(&head, q))
3275                         goto switch_back;
3276
3277         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3278                 blk_mq_debugfs_unregister_hctxs(q);
3279                 blk_mq_sysfs_unregister(q);
3280         }
3281
3282         prev_nr_hw_queues = set->nr_hw_queues;
3283         set->nr_hw_queues = nr_hw_queues;
3284         blk_mq_update_queue_map(set);
3285 fallback:
3286         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3287                 blk_mq_realloc_hw_ctxs(set, q);
3288                 if (q->nr_hw_queues != set->nr_hw_queues) {
3289                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3290                                         nr_hw_queues, prev_nr_hw_queues);
3291                         set->nr_hw_queues = prev_nr_hw_queues;
3292                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3293                         goto fallback;
3294                 }
3295                 blk_mq_map_swqueue(q);
3296         }
3297
3298         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3299                 blk_mq_sysfs_register(q);
3300                 blk_mq_debugfs_register_hctxs(q);
3301         }
3302
3303 switch_back:
3304         list_for_each_entry(q, &set->tag_list, tag_set_list)
3305                 blk_mq_elv_switch_back(&head, q);
3306
3307         list_for_each_entry(q, &set->tag_list, tag_set_list)
3308                 blk_mq_unfreeze_queue(q);
3309 }
3310
3311 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3312 {
3313         mutex_lock(&set->tag_list_lock);
3314         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3315         mutex_unlock(&set->tag_list_lock);
3316 }
3317 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3318
3319 /* Enable polling stats and return whether they were already enabled. */
3320 static bool blk_poll_stats_enable(struct request_queue *q)
3321 {
3322         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3323             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3324                 return true;
3325         blk_stat_add_callback(q, q->poll_cb);
3326         return false;
3327 }
3328
3329 static void blk_mq_poll_stats_start(struct request_queue *q)
3330 {
3331         /*
3332          * We don't arm the callback if polling stats are not enabled or the
3333          * callback is already active.
3334          */
3335         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3336             blk_stat_is_active(q->poll_cb))
3337                 return;
3338
3339         blk_stat_activate_msecs(q->poll_cb, 100);
3340 }
3341
3342 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3343 {
3344         struct request_queue *q = cb->data;
3345         int bucket;
3346
3347         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3348                 if (cb->stat[bucket].nr_samples)
3349                         q->poll_stat[bucket] = cb->stat[bucket];
3350         }
3351 }
3352
3353 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3354                                        struct blk_mq_hw_ctx *hctx,
3355                                        struct request *rq)
3356 {
3357         unsigned long ret = 0;
3358         int bucket;
3359
3360         /*
3361          * If stats collection isn't on, don't sleep but turn it on for
3362          * future users
3363          */
3364         if (!blk_poll_stats_enable(q))
3365                 return 0;
3366
3367         /*
3368          * As an optimistic guess, use half of the mean service time
3369          * for this type of request. We can (and should) make this smarter.
3370          * For instance, if the completion latencies are tight, we can
3371          * get closer than just half the mean. This is especially
3372          * important on devices where the completion latencies are longer
3373          * than ~10 usec. We do use the stats for the relevant IO size
3374          * if available which does lead to better estimates.
3375          */
3376         bucket = blk_mq_poll_stats_bkt(rq);
3377         if (bucket < 0)
3378                 return ret;
3379
3380         if (q->poll_stat[bucket].nr_samples)
3381                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3382
3383         return ret;
3384 }
3385
3386 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3387                                      struct blk_mq_hw_ctx *hctx,
3388                                      struct request *rq)
3389 {
3390         struct hrtimer_sleeper hs;
3391         enum hrtimer_mode mode;
3392         unsigned int nsecs;
3393         ktime_t kt;
3394
3395         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3396                 return false;
3397
3398         /*
3399          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3400          *
3401          *  0:  use half of prev avg
3402          * >0:  use this specific value
3403          */
3404         if (q->poll_nsec > 0)
3405                 nsecs = q->poll_nsec;
3406         else
3407                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3408
3409         if (!nsecs)
3410                 return false;
3411
3412         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3413
3414         /*
3415          * This will be replaced with the stats tracking code, using
3416          * 'avg_completion_time / 2' as the pre-sleep target.
3417          */
3418         kt = nsecs;
3419
3420         mode = HRTIMER_MODE_REL;
3421         hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3422         hrtimer_set_expires(&hs.timer, kt);
3423
3424         do {
3425                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3426                         break;
3427                 set_current_state(TASK_UNINTERRUPTIBLE);
3428                 hrtimer_sleeper_start_expires(&hs, mode);
3429                 if (hs.task)
3430                         io_schedule();
3431                 hrtimer_cancel(&hs.timer);
3432                 mode = HRTIMER_MODE_ABS;
3433         } while (hs.task && !signal_pending(current));
3434
3435         __set_current_state(TASK_RUNNING);
3436         destroy_hrtimer_on_stack(&hs.timer);
3437         return true;
3438 }
3439
3440 static bool blk_mq_poll_hybrid(struct request_queue *q,
3441                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3442 {
3443         struct request *rq;
3444
3445         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3446                 return false;
3447
3448         if (!blk_qc_t_is_internal(cookie))
3449                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3450         else {
3451                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3452                 /*
3453                  * With scheduling, if the request has completed, we'll
3454                  * get a NULL return here, as we clear the sched tag when
3455                  * that happens. The request still remains valid, like always,
3456                  * so we should be safe with just the NULL check.
3457                  */
3458                 if (!rq)
3459                         return false;
3460         }
3461
3462         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3463 }
3464
3465 /**
3466  * blk_poll - poll for IO completions
3467  * @q:  the queue
3468  * @cookie: cookie passed back at IO submission time
3469  * @spin: whether to spin for completions
3470  *
3471  * Description:
3472  *    Poll for completions on the passed in queue. Returns number of
3473  *    completed entries found. If @spin is true, then blk_poll will continue
3474  *    looping until at least one completion is found, unless the task is
3475  *    otherwise marked running (or we need to reschedule).
3476  */
3477 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3478 {
3479         struct blk_mq_hw_ctx *hctx;
3480         long state;
3481
3482         if (!blk_qc_t_valid(cookie) ||
3483             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3484                 return 0;
3485
3486         if (current->plug)
3487                 blk_flush_plug_list(current->plug, false);
3488
3489         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3490
3491         /*
3492          * If we sleep, have the caller restart the poll loop to reset
3493          * the state. Like for the other success return cases, the
3494          * caller is responsible for checking if the IO completed. If
3495          * the IO isn't complete, we'll get called again and will go
3496          * straight to the busy poll loop.
3497          */
3498         if (blk_mq_poll_hybrid(q, hctx, cookie))
3499                 return 1;
3500
3501         hctx->poll_considered++;
3502
3503         state = current->state;
3504         do {
3505                 int ret;
3506
3507                 hctx->poll_invoked++;
3508
3509                 ret = q->mq_ops->poll(hctx);
3510                 if (ret > 0) {
3511                         hctx->poll_success++;
3512                         __set_current_state(TASK_RUNNING);
3513                         return ret;
3514                 }
3515
3516                 if (signal_pending_state(state, current))
3517                         __set_current_state(TASK_RUNNING);
3518
3519                 if (current->state == TASK_RUNNING)
3520                         return 1;
3521                 if (ret < 0 || !spin)
3522                         break;
3523                 cpu_relax();
3524         } while (!need_resched());
3525
3526         __set_current_state(TASK_RUNNING);
3527         return 0;
3528 }
3529 EXPORT_SYMBOL_GPL(blk_poll);
3530
3531 unsigned int blk_mq_rq_cpu(struct request *rq)
3532 {
3533         return rq->mq_ctx->cpu;
3534 }
3535 EXPORT_SYMBOL(blk_mq_rq_cpu);
3536
3537 static int __init blk_mq_init(void)
3538 {
3539         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3540                                 blk_mq_hctx_notify_dead);
3541         return 0;
3542 }
3543 subsys_initcall(blk_mq_init);